1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
11
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
21 #include "md.h"
22 #include "raid10.h"
23 #include "raid0.h"
24 #include "md-bitmap.h"
25
26 /*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
36 *
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
45 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
49 *
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
57 * on a device):
58 * A B C D A B C D E
59 * ... ...
60 * D A B C E A B C D
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
65 */
66
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69 static int _enough(struct r10conf *conf, int previous, int ignore);
70 static int enough(struct r10conf *conf, int ignore);
71 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
72 int *skipped);
73 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74 static void end_reshape_write(struct bio *bio);
75 static void end_reshape(struct r10conf *conf);
76
77 #define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
79
80 #include "raid1-10.c"
81
82 /*
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
85 */
get_resync_r10bio(struct bio * bio)86 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
87 {
88 return get_resync_pages(bio)->raid_bio;
89 }
90
r10bio_pool_alloc(gfp_t gfp_flags,void * data)91 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
92 {
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
95
96 /* allocate a r10bio with room for raid_disks entries in the
97 * bios array */
98 return kzalloc(size, gfp_flags);
99 }
100
101 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102 /* amount of memory to reserve for resync requests */
103 #define RESYNC_WINDOW (1024*1024)
104 /* maximum number of concurrent requests, memory permitting */
105 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
108
109 /*
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
114 *
115 */
r10buf_pool_alloc(gfp_t gfp_flags,void * data)116 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
117 {
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
120 struct bio *bio;
121 int j;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
124
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
126 if (!r10_bio)
127 return NULL;
128
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
132 else
133 nalloc = 2; /* recovery */
134
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
137 nalloc_rp = nalloc;
138 else
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
141 if (!rps)
142 goto out_free_r10bio;
143
144 /*
145 * Allocate bios.
146 */
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
149 if (!bio)
150 goto out_free_bio;
151 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
152 r10_bio->devs[j].bio = bio;
153 if (!conf->have_replacement)
154 continue;
155 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
156 if (!bio)
157 goto out_free_bio;
158 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
159 r10_bio->devs[j].repl_bio = bio;
160 }
161 /*
162 * Allocate RESYNC_PAGES data pages and attach them
163 * where needed.
164 */
165 for (j = 0; j < nalloc; j++) {
166 struct bio *rbio = r10_bio->devs[j].repl_bio;
167 struct resync_pages *rp, *rp_repl;
168
169 rp = &rps[j];
170 if (rbio)
171 rp_repl = &rps[nalloc + j];
172
173 bio = r10_bio->devs[j].bio;
174
175 if (!j || test_bit(MD_RECOVERY_SYNC,
176 &conf->mddev->recovery)) {
177 if (resync_alloc_pages(rp, gfp_flags))
178 goto out_free_pages;
179 } else {
180 memcpy(rp, &rps[0], sizeof(*rp));
181 resync_get_all_pages(rp);
182 }
183
184 rp->raid_bio = r10_bio;
185 bio->bi_private = rp;
186 if (rbio) {
187 memcpy(rp_repl, rp, sizeof(*rp));
188 rbio->bi_private = rp_repl;
189 }
190 }
191
192 return r10_bio;
193
194 out_free_pages:
195 while (--j >= 0)
196 resync_free_pages(&rps[j]);
197
198 j = 0;
199 out_free_bio:
200 for ( ; j < nalloc; j++) {
201 if (r10_bio->devs[j].bio)
202 bio_uninit(r10_bio->devs[j].bio);
203 kfree(r10_bio->devs[j].bio);
204 if (r10_bio->devs[j].repl_bio)
205 bio_uninit(r10_bio->devs[j].repl_bio);
206 kfree(r10_bio->devs[j].repl_bio);
207 }
208 kfree(rps);
209 out_free_r10bio:
210 rbio_pool_free(r10_bio, conf);
211 return NULL;
212 }
213
r10buf_pool_free(void * __r10_bio,void * data)214 static void r10buf_pool_free(void *__r10_bio, void *data)
215 {
216 struct r10conf *conf = data;
217 struct r10bio *r10bio = __r10_bio;
218 int j;
219 struct resync_pages *rp = NULL;
220
221 for (j = conf->copies; j--; ) {
222 struct bio *bio = r10bio->devs[j].bio;
223
224 if (bio) {
225 rp = get_resync_pages(bio);
226 resync_free_pages(rp);
227 bio_uninit(bio);
228 kfree(bio);
229 }
230
231 bio = r10bio->devs[j].repl_bio;
232 if (bio) {
233 bio_uninit(bio);
234 kfree(bio);
235 }
236 }
237
238 /* resync pages array stored in the 1st bio's .bi_private */
239 kfree(rp);
240
241 rbio_pool_free(r10bio, conf);
242 }
243
put_all_bios(struct r10conf * conf,struct r10bio * r10_bio)244 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
245 {
246 int i;
247
248 for (i = 0; i < conf->geo.raid_disks; i++) {
249 struct bio **bio = & r10_bio->devs[i].bio;
250 if (!BIO_SPECIAL(*bio))
251 bio_put(*bio);
252 *bio = NULL;
253 bio = &r10_bio->devs[i].repl_bio;
254 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
255 bio_put(*bio);
256 *bio = NULL;
257 }
258 }
259
free_r10bio(struct r10bio * r10_bio)260 static void free_r10bio(struct r10bio *r10_bio)
261 {
262 struct r10conf *conf = r10_bio->mddev->private;
263
264 put_all_bios(conf, r10_bio);
265 mempool_free(r10_bio, &conf->r10bio_pool);
266 }
267
put_buf(struct r10bio * r10_bio)268 static void put_buf(struct r10bio *r10_bio)
269 {
270 struct r10conf *conf = r10_bio->mddev->private;
271
272 mempool_free(r10_bio, &conf->r10buf_pool);
273
274 lower_barrier(conf);
275 }
276
reschedule_retry(struct r10bio * r10_bio)277 static void reschedule_retry(struct r10bio *r10_bio)
278 {
279 unsigned long flags;
280 struct mddev *mddev = r10_bio->mddev;
281 struct r10conf *conf = mddev->private;
282
283 spin_lock_irqsave(&conf->device_lock, flags);
284 list_add(&r10_bio->retry_list, &conf->retry_list);
285 conf->nr_queued ++;
286 spin_unlock_irqrestore(&conf->device_lock, flags);
287
288 /* wake up frozen array... */
289 wake_up(&conf->wait_barrier);
290
291 md_wakeup_thread(mddev->thread);
292 }
293
294 /*
295 * raid_end_bio_io() is called when we have finished servicing a mirrored
296 * operation and are ready to return a success/failure code to the buffer
297 * cache layer.
298 */
raid_end_bio_io(struct r10bio * r10_bio)299 static void raid_end_bio_io(struct r10bio *r10_bio)
300 {
301 struct bio *bio = r10_bio->master_bio;
302 struct r10conf *conf = r10_bio->mddev->private;
303
304 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
305 bio->bi_status = BLK_STS_IOERR;
306
307 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
308 bio_end_io_acct(bio, r10_bio->start_time);
309 bio_endio(bio);
310 /*
311 * Wake up any possible resync thread that waits for the device
312 * to go idle.
313 */
314 allow_barrier(conf);
315
316 free_r10bio(r10_bio);
317 }
318
319 /*
320 * Update disk head position estimator based on IRQ completion info.
321 */
update_head_pos(int slot,struct r10bio * r10_bio)322 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
323 {
324 struct r10conf *conf = r10_bio->mddev->private;
325
326 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
327 r10_bio->devs[slot].addr + (r10_bio->sectors);
328 }
329
330 /*
331 * Find the disk number which triggered given bio
332 */
find_bio_disk(struct r10conf * conf,struct r10bio * r10_bio,struct bio * bio,int * slotp,int * replp)333 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
334 struct bio *bio, int *slotp, int *replp)
335 {
336 int slot;
337 int repl = 0;
338
339 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
340 if (r10_bio->devs[slot].bio == bio)
341 break;
342 if (r10_bio->devs[slot].repl_bio == bio) {
343 repl = 1;
344 break;
345 }
346 }
347
348 update_head_pos(slot, r10_bio);
349
350 if (slotp)
351 *slotp = slot;
352 if (replp)
353 *replp = repl;
354 return r10_bio->devs[slot].devnum;
355 }
356
raid10_end_read_request(struct bio * bio)357 static void raid10_end_read_request(struct bio *bio)
358 {
359 int uptodate = !bio->bi_status;
360 struct r10bio *r10_bio = bio->bi_private;
361 int slot;
362 struct md_rdev *rdev;
363 struct r10conf *conf = r10_bio->mddev->private;
364
365 slot = r10_bio->read_slot;
366 rdev = r10_bio->devs[slot].rdev;
367 /*
368 * this branch is our 'one mirror IO has finished' event handler:
369 */
370 update_head_pos(slot, r10_bio);
371
372 if (uptodate) {
373 /*
374 * Set R10BIO_Uptodate in our master bio, so that
375 * we will return a good error code to the higher
376 * levels even if IO on some other mirrored buffer fails.
377 *
378 * The 'master' represents the composite IO operation to
379 * user-side. So if something waits for IO, then it will
380 * wait for the 'master' bio.
381 */
382 set_bit(R10BIO_Uptodate, &r10_bio->state);
383 } else {
384 /* If all other devices that store this block have
385 * failed, we want to return the error upwards rather
386 * than fail the last device. Here we redefine
387 * "uptodate" to mean "Don't want to retry"
388 */
389 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
390 rdev->raid_disk))
391 uptodate = 1;
392 }
393 if (uptodate) {
394 raid_end_bio_io(r10_bio);
395 rdev_dec_pending(rdev, conf->mddev);
396 } else {
397 /*
398 * oops, read error - keep the refcount on the rdev
399 */
400 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
401 mdname(conf->mddev),
402 rdev->bdev,
403 (unsigned long long)r10_bio->sector);
404 set_bit(R10BIO_ReadError, &r10_bio->state);
405 reschedule_retry(r10_bio);
406 }
407 }
408
close_write(struct r10bio * r10_bio)409 static void close_write(struct r10bio *r10_bio)
410 {
411 /* clear the bitmap if all writes complete successfully */
412 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
413 r10_bio->sectors,
414 !test_bit(R10BIO_Degraded, &r10_bio->state),
415 0);
416 md_write_end(r10_bio->mddev);
417 }
418
one_write_done(struct r10bio * r10_bio)419 static void one_write_done(struct r10bio *r10_bio)
420 {
421 if (atomic_dec_and_test(&r10_bio->remaining)) {
422 if (test_bit(R10BIO_WriteError, &r10_bio->state))
423 reschedule_retry(r10_bio);
424 else {
425 close_write(r10_bio);
426 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
427 reschedule_retry(r10_bio);
428 else
429 raid_end_bio_io(r10_bio);
430 }
431 }
432 }
433
raid10_end_write_request(struct bio * bio)434 static void raid10_end_write_request(struct bio *bio)
435 {
436 struct r10bio *r10_bio = bio->bi_private;
437 int dev;
438 int dec_rdev = 1;
439 struct r10conf *conf = r10_bio->mddev->private;
440 int slot, repl;
441 struct md_rdev *rdev = NULL;
442 struct bio *to_put = NULL;
443 bool discard_error;
444
445 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
446
447 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
448
449 if (repl)
450 rdev = conf->mirrors[dev].replacement;
451 if (!rdev) {
452 smp_rmb();
453 repl = 0;
454 rdev = conf->mirrors[dev].rdev;
455 }
456 /*
457 * this branch is our 'one mirror IO has finished' event handler:
458 */
459 if (bio->bi_status && !discard_error) {
460 if (repl)
461 /* Never record new bad blocks to replacement,
462 * just fail it.
463 */
464 md_error(rdev->mddev, rdev);
465 else {
466 set_bit(WriteErrorSeen, &rdev->flags);
467 if (!test_and_set_bit(WantReplacement, &rdev->flags))
468 set_bit(MD_RECOVERY_NEEDED,
469 &rdev->mddev->recovery);
470
471 dec_rdev = 0;
472 if (test_bit(FailFast, &rdev->flags) &&
473 (bio->bi_opf & MD_FAILFAST)) {
474 md_error(rdev->mddev, rdev);
475 }
476
477 /*
478 * When the device is faulty, it is not necessary to
479 * handle write error.
480 */
481 if (!test_bit(Faulty, &rdev->flags))
482 set_bit(R10BIO_WriteError, &r10_bio->state);
483 else {
484 /* Fail the request */
485 set_bit(R10BIO_Degraded, &r10_bio->state);
486 r10_bio->devs[slot].bio = NULL;
487 to_put = bio;
488 dec_rdev = 1;
489 }
490 }
491 } else {
492 /*
493 * Set R10BIO_Uptodate in our master bio, so that
494 * we will return a good error code for to the higher
495 * levels even if IO on some other mirrored buffer fails.
496 *
497 * The 'master' represents the composite IO operation to
498 * user-side. So if something waits for IO, then it will
499 * wait for the 'master' bio.
500 */
501 sector_t first_bad;
502 int bad_sectors;
503
504 /*
505 * Do not set R10BIO_Uptodate if the current device is
506 * rebuilding or Faulty. This is because we cannot use
507 * such device for properly reading the data back (we could
508 * potentially use it, if the current write would have felt
509 * before rdev->recovery_offset, but for simplicity we don't
510 * check this here.
511 */
512 if (test_bit(In_sync, &rdev->flags) &&
513 !test_bit(Faulty, &rdev->flags))
514 set_bit(R10BIO_Uptodate, &r10_bio->state);
515
516 /* Maybe we can clear some bad blocks. */
517 if (is_badblock(rdev,
518 r10_bio->devs[slot].addr,
519 r10_bio->sectors,
520 &first_bad, &bad_sectors) && !discard_error) {
521 bio_put(bio);
522 if (repl)
523 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
524 else
525 r10_bio->devs[slot].bio = IO_MADE_GOOD;
526 dec_rdev = 0;
527 set_bit(R10BIO_MadeGood, &r10_bio->state);
528 }
529 }
530
531 /*
532 *
533 * Let's see if all mirrored write operations have finished
534 * already.
535 */
536 one_write_done(r10_bio);
537 if (dec_rdev)
538 rdev_dec_pending(rdev, conf->mddev);
539 if (to_put)
540 bio_put(to_put);
541 }
542
543 /*
544 * RAID10 layout manager
545 * As well as the chunksize and raid_disks count, there are two
546 * parameters: near_copies and far_copies.
547 * near_copies * far_copies must be <= raid_disks.
548 * Normally one of these will be 1.
549 * If both are 1, we get raid0.
550 * If near_copies == raid_disks, we get raid1.
551 *
552 * Chunks are laid out in raid0 style with near_copies copies of the
553 * first chunk, followed by near_copies copies of the next chunk and
554 * so on.
555 * If far_copies > 1, then after 1/far_copies of the array has been assigned
556 * as described above, we start again with a device offset of near_copies.
557 * So we effectively have another copy of the whole array further down all
558 * the drives, but with blocks on different drives.
559 * With this layout, and block is never stored twice on the one device.
560 *
561 * raid10_find_phys finds the sector offset of a given virtual sector
562 * on each device that it is on.
563 *
564 * raid10_find_virt does the reverse mapping, from a device and a
565 * sector offset to a virtual address
566 */
567
__raid10_find_phys(struct geom * geo,struct r10bio * r10bio)568 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
569 {
570 int n,f;
571 sector_t sector;
572 sector_t chunk;
573 sector_t stripe;
574 int dev;
575 int slot = 0;
576 int last_far_set_start, last_far_set_size;
577
578 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
579 last_far_set_start *= geo->far_set_size;
580
581 last_far_set_size = geo->far_set_size;
582 last_far_set_size += (geo->raid_disks % geo->far_set_size);
583
584 /* now calculate first sector/dev */
585 chunk = r10bio->sector >> geo->chunk_shift;
586 sector = r10bio->sector & geo->chunk_mask;
587
588 chunk *= geo->near_copies;
589 stripe = chunk;
590 dev = sector_div(stripe, geo->raid_disks);
591 if (geo->far_offset)
592 stripe *= geo->far_copies;
593
594 sector += stripe << geo->chunk_shift;
595
596 /* and calculate all the others */
597 for (n = 0; n < geo->near_copies; n++) {
598 int d = dev;
599 int set;
600 sector_t s = sector;
601 r10bio->devs[slot].devnum = d;
602 r10bio->devs[slot].addr = s;
603 slot++;
604
605 for (f = 1; f < geo->far_copies; f++) {
606 set = d / geo->far_set_size;
607 d += geo->near_copies;
608
609 if ((geo->raid_disks % geo->far_set_size) &&
610 (d > last_far_set_start)) {
611 d -= last_far_set_start;
612 d %= last_far_set_size;
613 d += last_far_set_start;
614 } else {
615 d %= geo->far_set_size;
616 d += geo->far_set_size * set;
617 }
618 s += geo->stride;
619 r10bio->devs[slot].devnum = d;
620 r10bio->devs[slot].addr = s;
621 slot++;
622 }
623 dev++;
624 if (dev >= geo->raid_disks) {
625 dev = 0;
626 sector += (geo->chunk_mask + 1);
627 }
628 }
629 }
630
raid10_find_phys(struct r10conf * conf,struct r10bio * r10bio)631 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
632 {
633 struct geom *geo = &conf->geo;
634
635 if (conf->reshape_progress != MaxSector &&
636 ((r10bio->sector >= conf->reshape_progress) !=
637 conf->mddev->reshape_backwards)) {
638 set_bit(R10BIO_Previous, &r10bio->state);
639 geo = &conf->prev;
640 } else
641 clear_bit(R10BIO_Previous, &r10bio->state);
642
643 __raid10_find_phys(geo, r10bio);
644 }
645
raid10_find_virt(struct r10conf * conf,sector_t sector,int dev)646 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
647 {
648 sector_t offset, chunk, vchunk;
649 /* Never use conf->prev as this is only called during resync
650 * or recovery, so reshape isn't happening
651 */
652 struct geom *geo = &conf->geo;
653 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
654 int far_set_size = geo->far_set_size;
655 int last_far_set_start;
656
657 if (geo->raid_disks % geo->far_set_size) {
658 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
659 last_far_set_start *= geo->far_set_size;
660
661 if (dev >= last_far_set_start) {
662 far_set_size = geo->far_set_size;
663 far_set_size += (geo->raid_disks % geo->far_set_size);
664 far_set_start = last_far_set_start;
665 }
666 }
667
668 offset = sector & geo->chunk_mask;
669 if (geo->far_offset) {
670 int fc;
671 chunk = sector >> geo->chunk_shift;
672 fc = sector_div(chunk, geo->far_copies);
673 dev -= fc * geo->near_copies;
674 if (dev < far_set_start)
675 dev += far_set_size;
676 } else {
677 while (sector >= geo->stride) {
678 sector -= geo->stride;
679 if (dev < (geo->near_copies + far_set_start))
680 dev += far_set_size - geo->near_copies;
681 else
682 dev -= geo->near_copies;
683 }
684 chunk = sector >> geo->chunk_shift;
685 }
686 vchunk = chunk * geo->raid_disks + dev;
687 sector_div(vchunk, geo->near_copies);
688 return (vchunk << geo->chunk_shift) + offset;
689 }
690
691 /*
692 * This routine returns the disk from which the requested read should
693 * be done. There is a per-array 'next expected sequential IO' sector
694 * number - if this matches on the next IO then we use the last disk.
695 * There is also a per-disk 'last know head position' sector that is
696 * maintained from IRQ contexts, both the normal and the resync IO
697 * completion handlers update this position correctly. If there is no
698 * perfect sequential match then we pick the disk whose head is closest.
699 *
700 * If there are 2 mirrors in the same 2 devices, performance degrades
701 * because position is mirror, not device based.
702 *
703 * The rdev for the device selected will have nr_pending incremented.
704 */
705
706 /*
707 * FIXME: possibly should rethink readbalancing and do it differently
708 * depending on near_copies / far_copies geometry.
709 */
read_balance(struct r10conf * conf,struct r10bio * r10_bio,int * max_sectors)710 static struct md_rdev *read_balance(struct r10conf *conf,
711 struct r10bio *r10_bio,
712 int *max_sectors)
713 {
714 const sector_t this_sector = r10_bio->sector;
715 int disk, slot;
716 int sectors = r10_bio->sectors;
717 int best_good_sectors;
718 sector_t new_distance, best_dist;
719 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
720 int do_balance;
721 int best_dist_slot, best_pending_slot;
722 bool has_nonrot_disk = false;
723 unsigned int min_pending;
724 struct geom *geo = &conf->geo;
725
726 raid10_find_phys(conf, r10_bio);
727 rcu_read_lock();
728 best_dist_slot = -1;
729 min_pending = UINT_MAX;
730 best_dist_rdev = NULL;
731 best_pending_rdev = NULL;
732 best_dist = MaxSector;
733 best_good_sectors = 0;
734 do_balance = 1;
735 clear_bit(R10BIO_FailFast, &r10_bio->state);
736 /*
737 * Check if we can balance. We can balance on the whole
738 * device if no resync is going on (recovery is ok), or below
739 * the resync window. We take the first readable disk when
740 * above the resync window.
741 */
742 if ((conf->mddev->recovery_cp < MaxSector
743 && (this_sector + sectors >= conf->next_resync)) ||
744 (mddev_is_clustered(conf->mddev) &&
745 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
746 this_sector + sectors)))
747 do_balance = 0;
748
749 for (slot = 0; slot < conf->copies ; slot++) {
750 sector_t first_bad;
751 int bad_sectors;
752 sector_t dev_sector;
753 unsigned int pending;
754 bool nonrot;
755
756 if (r10_bio->devs[slot].bio == IO_BLOCKED)
757 continue;
758 disk = r10_bio->devs[slot].devnum;
759 rdev = rcu_dereference(conf->mirrors[disk].replacement);
760 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
761 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
762 rdev = rcu_dereference(conf->mirrors[disk].rdev);
763 if (rdev == NULL ||
764 test_bit(Faulty, &rdev->flags))
765 continue;
766 if (!test_bit(In_sync, &rdev->flags) &&
767 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
768 continue;
769
770 dev_sector = r10_bio->devs[slot].addr;
771 if (is_badblock(rdev, dev_sector, sectors,
772 &first_bad, &bad_sectors)) {
773 if (best_dist < MaxSector)
774 /* Already have a better slot */
775 continue;
776 if (first_bad <= dev_sector) {
777 /* Cannot read here. If this is the
778 * 'primary' device, then we must not read
779 * beyond 'bad_sectors' from another device.
780 */
781 bad_sectors -= (dev_sector - first_bad);
782 if (!do_balance && sectors > bad_sectors)
783 sectors = bad_sectors;
784 if (best_good_sectors > sectors)
785 best_good_sectors = sectors;
786 } else {
787 sector_t good_sectors =
788 first_bad - dev_sector;
789 if (good_sectors > best_good_sectors) {
790 best_good_sectors = good_sectors;
791 best_dist_slot = slot;
792 best_dist_rdev = rdev;
793 }
794 if (!do_balance)
795 /* Must read from here */
796 break;
797 }
798 continue;
799 } else
800 best_good_sectors = sectors;
801
802 if (!do_balance)
803 break;
804
805 nonrot = bdev_nonrot(rdev->bdev);
806 has_nonrot_disk |= nonrot;
807 pending = atomic_read(&rdev->nr_pending);
808 if (min_pending > pending && nonrot) {
809 min_pending = pending;
810 best_pending_slot = slot;
811 best_pending_rdev = rdev;
812 }
813
814 if (best_dist_slot >= 0)
815 /* At least 2 disks to choose from so failfast is OK */
816 set_bit(R10BIO_FailFast, &r10_bio->state);
817 /* This optimisation is debatable, and completely destroys
818 * sequential read speed for 'far copies' arrays. So only
819 * keep it for 'near' arrays, and review those later.
820 */
821 if (geo->near_copies > 1 && !pending)
822 new_distance = 0;
823
824 /* for far > 1 always use the lowest address */
825 else if (geo->far_copies > 1)
826 new_distance = r10_bio->devs[slot].addr;
827 else
828 new_distance = abs(r10_bio->devs[slot].addr -
829 conf->mirrors[disk].head_position);
830
831 if (new_distance < best_dist) {
832 best_dist = new_distance;
833 best_dist_slot = slot;
834 best_dist_rdev = rdev;
835 }
836 }
837 if (slot >= conf->copies) {
838 if (has_nonrot_disk) {
839 slot = best_pending_slot;
840 rdev = best_pending_rdev;
841 } else {
842 slot = best_dist_slot;
843 rdev = best_dist_rdev;
844 }
845 }
846
847 if (slot >= 0) {
848 atomic_inc(&rdev->nr_pending);
849 r10_bio->read_slot = slot;
850 } else
851 rdev = NULL;
852 rcu_read_unlock();
853 *max_sectors = best_good_sectors;
854
855 return rdev;
856 }
857
flush_pending_writes(struct r10conf * conf)858 static void flush_pending_writes(struct r10conf *conf)
859 {
860 /* Any writes that have been queued but are awaiting
861 * bitmap updates get flushed here.
862 */
863 spin_lock_irq(&conf->device_lock);
864
865 if (conf->pending_bio_list.head) {
866 struct blk_plug plug;
867 struct bio *bio;
868
869 bio = bio_list_get(&conf->pending_bio_list);
870 spin_unlock_irq(&conf->device_lock);
871
872 /*
873 * As this is called in a wait_event() loop (see freeze_array),
874 * current->state might be TASK_UNINTERRUPTIBLE which will
875 * cause a warning when we prepare to wait again. As it is
876 * rare that this path is taken, it is perfectly safe to force
877 * us to go around the wait_event() loop again, so the warning
878 * is a false-positive. Silence the warning by resetting
879 * thread state
880 */
881 __set_current_state(TASK_RUNNING);
882
883 blk_start_plug(&plug);
884 /* flush any pending bitmap writes to disk
885 * before proceeding w/ I/O */
886 md_bitmap_unplug(conf->mddev->bitmap);
887 wake_up(&conf->wait_barrier);
888
889 while (bio) { /* submit pending writes */
890 struct bio *next = bio->bi_next;
891 struct md_rdev *rdev = (void*)bio->bi_bdev;
892 bio->bi_next = NULL;
893 bio_set_dev(bio, rdev->bdev);
894 if (test_bit(Faulty, &rdev->flags)) {
895 bio_io_error(bio);
896 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
897 !bdev_max_discard_sectors(bio->bi_bdev)))
898 /* Just ignore it */
899 bio_endio(bio);
900 else
901 submit_bio_noacct(bio);
902 bio = next;
903 }
904 blk_finish_plug(&plug);
905 } else
906 spin_unlock_irq(&conf->device_lock);
907 }
908
909 /* Barriers....
910 * Sometimes we need to suspend IO while we do something else,
911 * either some resync/recovery, or reconfigure the array.
912 * To do this we raise a 'barrier'.
913 * The 'barrier' is a counter that can be raised multiple times
914 * to count how many activities are happening which preclude
915 * normal IO.
916 * We can only raise the barrier if there is no pending IO.
917 * i.e. if nr_pending == 0.
918 * We choose only to raise the barrier if no-one is waiting for the
919 * barrier to go down. This means that as soon as an IO request
920 * is ready, no other operations which require a barrier will start
921 * until the IO request has had a chance.
922 *
923 * So: regular IO calls 'wait_barrier'. When that returns there
924 * is no backgroup IO happening, It must arrange to call
925 * allow_barrier when it has finished its IO.
926 * backgroup IO calls must call raise_barrier. Once that returns
927 * there is no normal IO happeing. It must arrange to call
928 * lower_barrier when the particular background IO completes.
929 */
930
raise_barrier(struct r10conf * conf,int force)931 static void raise_barrier(struct r10conf *conf, int force)
932 {
933 BUG_ON(force && !conf->barrier);
934 spin_lock_irq(&conf->resync_lock);
935
936 /* Wait until no block IO is waiting (unless 'force') */
937 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
938 conf->resync_lock);
939
940 /* block any new IO from starting */
941 conf->barrier++;
942
943 /* Now wait for all pending IO to complete */
944 wait_event_lock_irq(conf->wait_barrier,
945 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
946 conf->resync_lock);
947
948 spin_unlock_irq(&conf->resync_lock);
949 }
950
lower_barrier(struct r10conf * conf)951 static void lower_barrier(struct r10conf *conf)
952 {
953 unsigned long flags;
954 spin_lock_irqsave(&conf->resync_lock, flags);
955 conf->barrier--;
956 spin_unlock_irqrestore(&conf->resync_lock, flags);
957 wake_up(&conf->wait_barrier);
958 }
959
wait_barrier(struct r10conf * conf,bool nowait)960 static bool wait_barrier(struct r10conf *conf, bool nowait)
961 {
962 bool ret = true;
963
964 spin_lock_irq(&conf->resync_lock);
965 if (conf->barrier) {
966 struct bio_list *bio_list = current->bio_list;
967 conf->nr_waiting++;
968 /* Wait for the barrier to drop.
969 * However if there are already pending
970 * requests (preventing the barrier from
971 * rising completely), and the
972 * pre-process bio queue isn't empty,
973 * then don't wait, as we need to empty
974 * that queue to get the nr_pending
975 * count down.
976 */
977 /* Return false when nowait flag is set */
978 if (nowait) {
979 ret = false;
980 } else {
981 raid10_log(conf->mddev, "wait barrier");
982 wait_event_lock_irq(conf->wait_barrier,
983 !conf->barrier ||
984 (atomic_read(&conf->nr_pending) &&
985 bio_list &&
986 (!bio_list_empty(&bio_list[0]) ||
987 !bio_list_empty(&bio_list[1]))) ||
988 /* move on if recovery thread is
989 * blocked by us
990 */
991 (conf->mddev->thread->tsk == current &&
992 test_bit(MD_RECOVERY_RUNNING,
993 &conf->mddev->recovery) &&
994 conf->nr_queued > 0),
995 conf->resync_lock);
996 }
997 conf->nr_waiting--;
998 if (!conf->nr_waiting)
999 wake_up(&conf->wait_barrier);
1000 }
1001 /* Only increment nr_pending when we wait */
1002 if (ret)
1003 atomic_inc(&conf->nr_pending);
1004 spin_unlock_irq(&conf->resync_lock);
1005 return ret;
1006 }
1007
allow_barrier(struct r10conf * conf)1008 static void allow_barrier(struct r10conf *conf)
1009 {
1010 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1011 (conf->array_freeze_pending))
1012 wake_up(&conf->wait_barrier);
1013 }
1014
freeze_array(struct r10conf * conf,int extra)1015 static void freeze_array(struct r10conf *conf, int extra)
1016 {
1017 /* stop syncio and normal IO and wait for everything to
1018 * go quiet.
1019 * We increment barrier and nr_waiting, and then
1020 * wait until nr_pending match nr_queued+extra
1021 * This is called in the context of one normal IO request
1022 * that has failed. Thus any sync request that might be pending
1023 * will be blocked by nr_pending, and we need to wait for
1024 * pending IO requests to complete or be queued for re-try.
1025 * Thus the number queued (nr_queued) plus this request (extra)
1026 * must match the number of pending IOs (nr_pending) before
1027 * we continue.
1028 */
1029 spin_lock_irq(&conf->resync_lock);
1030 conf->array_freeze_pending++;
1031 conf->barrier++;
1032 conf->nr_waiting++;
1033 wait_event_lock_irq_cmd(conf->wait_barrier,
1034 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1035 conf->resync_lock,
1036 flush_pending_writes(conf));
1037
1038 conf->array_freeze_pending--;
1039 spin_unlock_irq(&conf->resync_lock);
1040 }
1041
unfreeze_array(struct r10conf * conf)1042 static void unfreeze_array(struct r10conf *conf)
1043 {
1044 /* reverse the effect of the freeze */
1045 spin_lock_irq(&conf->resync_lock);
1046 conf->barrier--;
1047 conf->nr_waiting--;
1048 wake_up(&conf->wait_barrier);
1049 spin_unlock_irq(&conf->resync_lock);
1050 }
1051
choose_data_offset(struct r10bio * r10_bio,struct md_rdev * rdev)1052 static sector_t choose_data_offset(struct r10bio *r10_bio,
1053 struct md_rdev *rdev)
1054 {
1055 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1056 test_bit(R10BIO_Previous, &r10_bio->state))
1057 return rdev->data_offset;
1058 else
1059 return rdev->new_data_offset;
1060 }
1061
raid10_unplug(struct blk_plug_cb * cb,bool from_schedule)1062 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1063 {
1064 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1065 struct mddev *mddev = plug->cb.data;
1066 struct r10conf *conf = mddev->private;
1067 struct bio *bio;
1068
1069 if (from_schedule || current->bio_list) {
1070 spin_lock_irq(&conf->device_lock);
1071 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1072 spin_unlock_irq(&conf->device_lock);
1073 wake_up(&conf->wait_barrier);
1074 md_wakeup_thread(mddev->thread);
1075 kfree(plug);
1076 return;
1077 }
1078
1079 /* we aren't scheduling, so we can do the write-out directly. */
1080 bio = bio_list_get(&plug->pending);
1081 md_bitmap_unplug(mddev->bitmap);
1082 wake_up(&conf->wait_barrier);
1083
1084 while (bio) { /* submit pending writes */
1085 struct bio *next = bio->bi_next;
1086 struct md_rdev *rdev = (void*)bio->bi_bdev;
1087 bio->bi_next = NULL;
1088 bio_set_dev(bio, rdev->bdev);
1089 if (test_bit(Faulty, &rdev->flags)) {
1090 bio_io_error(bio);
1091 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1092 !bdev_max_discard_sectors(bio->bi_bdev)))
1093 /* Just ignore it */
1094 bio_endio(bio);
1095 else
1096 submit_bio_noacct(bio);
1097 bio = next;
1098 }
1099 kfree(plug);
1100 }
1101
1102 /*
1103 * 1. Register the new request and wait if the reconstruction thread has put
1104 * up a bar for new requests. Continue immediately if no resync is active
1105 * currently.
1106 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1107 */
regular_request_wait(struct mddev * mddev,struct r10conf * conf,struct bio * bio,sector_t sectors)1108 static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1109 struct bio *bio, sector_t sectors)
1110 {
1111 /* Bail out if REQ_NOWAIT is set for the bio */
1112 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1113 bio_wouldblock_error(bio);
1114 return false;
1115 }
1116 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1117 bio->bi_iter.bi_sector < conf->reshape_progress &&
1118 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1119 allow_barrier(conf);
1120 if (bio->bi_opf & REQ_NOWAIT) {
1121 bio_wouldblock_error(bio);
1122 return false;
1123 }
1124 raid10_log(conf->mddev, "wait reshape");
1125 wait_event(conf->wait_barrier,
1126 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1127 conf->reshape_progress >= bio->bi_iter.bi_sector +
1128 sectors);
1129 wait_barrier(conf, false);
1130 }
1131 return true;
1132 }
1133
raid10_read_request(struct mddev * mddev,struct bio * bio,struct r10bio * r10_bio)1134 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1135 struct r10bio *r10_bio)
1136 {
1137 struct r10conf *conf = mddev->private;
1138 struct bio *read_bio;
1139 const int op = bio_op(bio);
1140 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1141 int max_sectors;
1142 struct md_rdev *rdev;
1143 char b[BDEVNAME_SIZE];
1144 int slot = r10_bio->read_slot;
1145 struct md_rdev *err_rdev = NULL;
1146 gfp_t gfp = GFP_NOIO;
1147
1148 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1149 /*
1150 * This is an error retry, but we cannot
1151 * safely dereference the rdev in the r10_bio,
1152 * we must use the one in conf.
1153 * If it has already been disconnected (unlikely)
1154 * we lose the device name in error messages.
1155 */
1156 int disk;
1157 /*
1158 * As we are blocking raid10, it is a little safer to
1159 * use __GFP_HIGH.
1160 */
1161 gfp = GFP_NOIO | __GFP_HIGH;
1162
1163 rcu_read_lock();
1164 disk = r10_bio->devs[slot].devnum;
1165 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1166 if (err_rdev)
1167 bdevname(err_rdev->bdev, b);
1168 else {
1169 strcpy(b, "???");
1170 /* This never gets dereferenced */
1171 err_rdev = r10_bio->devs[slot].rdev;
1172 }
1173 rcu_read_unlock();
1174 }
1175
1176 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1177 return;
1178 rdev = read_balance(conf, r10_bio, &max_sectors);
1179 if (!rdev) {
1180 if (err_rdev) {
1181 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1182 mdname(mddev), b,
1183 (unsigned long long)r10_bio->sector);
1184 }
1185 raid_end_bio_io(r10_bio);
1186 return;
1187 }
1188 if (err_rdev)
1189 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1190 mdname(mddev),
1191 rdev->bdev,
1192 (unsigned long long)r10_bio->sector);
1193 if (max_sectors < bio_sectors(bio)) {
1194 struct bio *split = bio_split(bio, max_sectors,
1195 gfp, &conf->bio_split);
1196 bio_chain(split, bio);
1197 allow_barrier(conf);
1198 submit_bio_noacct(bio);
1199 wait_barrier(conf, false);
1200 bio = split;
1201 r10_bio->master_bio = bio;
1202 r10_bio->sectors = max_sectors;
1203 }
1204 slot = r10_bio->read_slot;
1205
1206 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1207 r10_bio->start_time = bio_start_io_acct(bio);
1208 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1209
1210 r10_bio->devs[slot].bio = read_bio;
1211 r10_bio->devs[slot].rdev = rdev;
1212
1213 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1214 choose_data_offset(r10_bio, rdev);
1215 read_bio->bi_end_io = raid10_end_read_request;
1216 bio_set_op_attrs(read_bio, op, do_sync);
1217 if (test_bit(FailFast, &rdev->flags) &&
1218 test_bit(R10BIO_FailFast, &r10_bio->state))
1219 read_bio->bi_opf |= MD_FAILFAST;
1220 read_bio->bi_private = r10_bio;
1221
1222 if (mddev->gendisk)
1223 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1224 r10_bio->sector);
1225 submit_bio_noacct(read_bio);
1226 return;
1227 }
1228
raid10_write_one_disk(struct mddev * mddev,struct r10bio * r10_bio,struct bio * bio,bool replacement,int n_copy)1229 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1230 struct bio *bio, bool replacement,
1231 int n_copy)
1232 {
1233 const int op = bio_op(bio);
1234 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1235 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1236 unsigned long flags;
1237 struct blk_plug_cb *cb;
1238 struct raid1_plug_cb *plug = NULL;
1239 struct r10conf *conf = mddev->private;
1240 struct md_rdev *rdev;
1241 int devnum = r10_bio->devs[n_copy].devnum;
1242 struct bio *mbio;
1243
1244 if (replacement) {
1245 rdev = conf->mirrors[devnum].replacement;
1246 if (rdev == NULL) {
1247 /* Replacement just got moved to main 'rdev' */
1248 smp_mb();
1249 rdev = conf->mirrors[devnum].rdev;
1250 }
1251 } else
1252 rdev = conf->mirrors[devnum].rdev;
1253
1254 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1255 if (replacement)
1256 r10_bio->devs[n_copy].repl_bio = mbio;
1257 else
1258 r10_bio->devs[n_copy].bio = mbio;
1259
1260 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1261 choose_data_offset(r10_bio, rdev));
1262 mbio->bi_end_io = raid10_end_write_request;
1263 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1264 if (!replacement && test_bit(FailFast,
1265 &conf->mirrors[devnum].rdev->flags)
1266 && enough(conf, devnum))
1267 mbio->bi_opf |= MD_FAILFAST;
1268 mbio->bi_private = r10_bio;
1269
1270 if (conf->mddev->gendisk)
1271 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1272 r10_bio->sector);
1273 /* flush_pending_writes() needs access to the rdev so...*/
1274 mbio->bi_bdev = (void *)rdev;
1275
1276 atomic_inc(&r10_bio->remaining);
1277
1278 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1279 if (cb)
1280 plug = container_of(cb, struct raid1_plug_cb, cb);
1281 else
1282 plug = NULL;
1283 if (plug) {
1284 bio_list_add(&plug->pending, mbio);
1285 } else {
1286 spin_lock_irqsave(&conf->device_lock, flags);
1287 bio_list_add(&conf->pending_bio_list, mbio);
1288 spin_unlock_irqrestore(&conf->device_lock, flags);
1289 md_wakeup_thread(mddev->thread);
1290 }
1291 }
1292
wait_blocked_dev(struct mddev * mddev,struct r10bio * r10_bio)1293 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1294 {
1295 int i;
1296 struct r10conf *conf = mddev->private;
1297 struct md_rdev *blocked_rdev;
1298
1299 retry_wait:
1300 blocked_rdev = NULL;
1301 rcu_read_lock();
1302 for (i = 0; i < conf->copies; i++) {
1303 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1304 struct md_rdev *rrdev = rcu_dereference(
1305 conf->mirrors[i].replacement);
1306 if (rdev == rrdev)
1307 rrdev = NULL;
1308 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1309 atomic_inc(&rdev->nr_pending);
1310 blocked_rdev = rdev;
1311 break;
1312 }
1313 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1314 atomic_inc(&rrdev->nr_pending);
1315 blocked_rdev = rrdev;
1316 break;
1317 }
1318
1319 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1320 sector_t first_bad;
1321 sector_t dev_sector = r10_bio->devs[i].addr;
1322 int bad_sectors;
1323 int is_bad;
1324
1325 /*
1326 * Discard request doesn't care the write result
1327 * so it doesn't need to wait blocked disk here.
1328 */
1329 if (!r10_bio->sectors)
1330 continue;
1331
1332 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1333 &first_bad, &bad_sectors);
1334 if (is_bad < 0) {
1335 /*
1336 * Mustn't write here until the bad block
1337 * is acknowledged
1338 */
1339 atomic_inc(&rdev->nr_pending);
1340 set_bit(BlockedBadBlocks, &rdev->flags);
1341 blocked_rdev = rdev;
1342 break;
1343 }
1344 }
1345 }
1346 rcu_read_unlock();
1347
1348 if (unlikely(blocked_rdev)) {
1349 /* Have to wait for this device to get unblocked, then retry */
1350 allow_barrier(conf);
1351 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1352 __func__, blocked_rdev->raid_disk);
1353 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1354 wait_barrier(conf, false);
1355 goto retry_wait;
1356 }
1357 }
1358
raid10_write_request(struct mddev * mddev,struct bio * bio,struct r10bio * r10_bio)1359 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1360 struct r10bio *r10_bio)
1361 {
1362 struct r10conf *conf = mddev->private;
1363 int i;
1364 sector_t sectors;
1365 int max_sectors;
1366
1367 if ((mddev_is_clustered(mddev) &&
1368 md_cluster_ops->area_resyncing(mddev, WRITE,
1369 bio->bi_iter.bi_sector,
1370 bio_end_sector(bio)))) {
1371 DEFINE_WAIT(w);
1372 /* Bail out if REQ_NOWAIT is set for the bio */
1373 if (bio->bi_opf & REQ_NOWAIT) {
1374 bio_wouldblock_error(bio);
1375 return;
1376 }
1377 for (;;) {
1378 prepare_to_wait(&conf->wait_barrier,
1379 &w, TASK_IDLE);
1380 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1381 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1382 break;
1383 schedule();
1384 }
1385 finish_wait(&conf->wait_barrier, &w);
1386 }
1387
1388 sectors = r10_bio->sectors;
1389 if (!regular_request_wait(mddev, conf, bio, sectors))
1390 return;
1391 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1392 (mddev->reshape_backwards
1393 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1394 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1395 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1396 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1397 /* Need to update reshape_position in metadata */
1398 mddev->reshape_position = conf->reshape_progress;
1399 set_mask_bits(&mddev->sb_flags, 0,
1400 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1401 md_wakeup_thread(mddev->thread);
1402 if (bio->bi_opf & REQ_NOWAIT) {
1403 allow_barrier(conf);
1404 bio_wouldblock_error(bio);
1405 return;
1406 }
1407 raid10_log(conf->mddev, "wait reshape metadata");
1408 wait_event(mddev->sb_wait,
1409 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1410
1411 conf->reshape_safe = mddev->reshape_position;
1412 }
1413
1414 /* first select target devices under rcu_lock and
1415 * inc refcount on their rdev. Record them by setting
1416 * bios[x] to bio
1417 * If there are known/acknowledged bad blocks on any device
1418 * on which we have seen a write error, we want to avoid
1419 * writing to those blocks. This potentially requires several
1420 * writes to write around the bad blocks. Each set of writes
1421 * gets its own r10_bio with a set of bios attached.
1422 */
1423
1424 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1425 raid10_find_phys(conf, r10_bio);
1426
1427 wait_blocked_dev(mddev, r10_bio);
1428
1429 rcu_read_lock();
1430 max_sectors = r10_bio->sectors;
1431
1432 for (i = 0; i < conf->copies; i++) {
1433 int d = r10_bio->devs[i].devnum;
1434 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1435 struct md_rdev *rrdev = rcu_dereference(
1436 conf->mirrors[d].replacement);
1437 if (rdev == rrdev)
1438 rrdev = NULL;
1439 if (rdev && (test_bit(Faulty, &rdev->flags)))
1440 rdev = NULL;
1441 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1442 rrdev = NULL;
1443
1444 r10_bio->devs[i].bio = NULL;
1445 r10_bio->devs[i].repl_bio = NULL;
1446
1447 if (!rdev && !rrdev) {
1448 set_bit(R10BIO_Degraded, &r10_bio->state);
1449 continue;
1450 }
1451 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1452 sector_t first_bad;
1453 sector_t dev_sector = r10_bio->devs[i].addr;
1454 int bad_sectors;
1455 int is_bad;
1456
1457 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1458 &first_bad, &bad_sectors);
1459 if (is_bad && first_bad <= dev_sector) {
1460 /* Cannot write here at all */
1461 bad_sectors -= (dev_sector - first_bad);
1462 if (bad_sectors < max_sectors)
1463 /* Mustn't write more than bad_sectors
1464 * to other devices yet
1465 */
1466 max_sectors = bad_sectors;
1467 /* We don't set R10BIO_Degraded as that
1468 * only applies if the disk is missing,
1469 * so it might be re-added, and we want to
1470 * know to recover this chunk.
1471 * In this case the device is here, and the
1472 * fact that this chunk is not in-sync is
1473 * recorded in the bad block log.
1474 */
1475 continue;
1476 }
1477 if (is_bad) {
1478 int good_sectors = first_bad - dev_sector;
1479 if (good_sectors < max_sectors)
1480 max_sectors = good_sectors;
1481 }
1482 }
1483 if (rdev) {
1484 r10_bio->devs[i].bio = bio;
1485 atomic_inc(&rdev->nr_pending);
1486 }
1487 if (rrdev) {
1488 r10_bio->devs[i].repl_bio = bio;
1489 atomic_inc(&rrdev->nr_pending);
1490 }
1491 }
1492 rcu_read_unlock();
1493
1494 if (max_sectors < r10_bio->sectors)
1495 r10_bio->sectors = max_sectors;
1496
1497 if (r10_bio->sectors < bio_sectors(bio)) {
1498 struct bio *split = bio_split(bio, r10_bio->sectors,
1499 GFP_NOIO, &conf->bio_split);
1500 bio_chain(split, bio);
1501 allow_barrier(conf);
1502 submit_bio_noacct(bio);
1503 wait_barrier(conf, false);
1504 bio = split;
1505 r10_bio->master_bio = bio;
1506 }
1507
1508 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1509 r10_bio->start_time = bio_start_io_acct(bio);
1510 atomic_set(&r10_bio->remaining, 1);
1511 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1512
1513 for (i = 0; i < conf->copies; i++) {
1514 if (r10_bio->devs[i].bio)
1515 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1516 if (r10_bio->devs[i].repl_bio)
1517 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1518 }
1519 one_write_done(r10_bio);
1520 }
1521
__make_request(struct mddev * mddev,struct bio * bio,int sectors)1522 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1523 {
1524 struct r10conf *conf = mddev->private;
1525 struct r10bio *r10_bio;
1526
1527 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1528
1529 r10_bio->master_bio = bio;
1530 r10_bio->sectors = sectors;
1531
1532 r10_bio->mddev = mddev;
1533 r10_bio->sector = bio->bi_iter.bi_sector;
1534 r10_bio->state = 0;
1535 r10_bio->read_slot = -1;
1536 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1537 conf->geo.raid_disks);
1538
1539 if (bio_data_dir(bio) == READ)
1540 raid10_read_request(mddev, bio, r10_bio);
1541 else
1542 raid10_write_request(mddev, bio, r10_bio);
1543 }
1544
raid_end_discard_bio(struct r10bio * r10bio)1545 static void raid_end_discard_bio(struct r10bio *r10bio)
1546 {
1547 struct r10conf *conf = r10bio->mddev->private;
1548 struct r10bio *first_r10bio;
1549
1550 while (atomic_dec_and_test(&r10bio->remaining)) {
1551
1552 allow_barrier(conf);
1553
1554 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1555 first_r10bio = (struct r10bio *)r10bio->master_bio;
1556 free_r10bio(r10bio);
1557 r10bio = first_r10bio;
1558 } else {
1559 md_write_end(r10bio->mddev);
1560 bio_endio(r10bio->master_bio);
1561 free_r10bio(r10bio);
1562 break;
1563 }
1564 }
1565 }
1566
raid10_end_discard_request(struct bio * bio)1567 static void raid10_end_discard_request(struct bio *bio)
1568 {
1569 struct r10bio *r10_bio = bio->bi_private;
1570 struct r10conf *conf = r10_bio->mddev->private;
1571 struct md_rdev *rdev = NULL;
1572 int dev;
1573 int slot, repl;
1574
1575 /*
1576 * We don't care the return value of discard bio
1577 */
1578 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1579 set_bit(R10BIO_Uptodate, &r10_bio->state);
1580
1581 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1582 if (repl)
1583 rdev = conf->mirrors[dev].replacement;
1584 if (!rdev) {
1585 /*
1586 * raid10_remove_disk uses smp_mb to make sure rdev is set to
1587 * replacement before setting replacement to NULL. It can read
1588 * rdev first without barrier protect even replacment is NULL
1589 */
1590 smp_rmb();
1591 rdev = conf->mirrors[dev].rdev;
1592 }
1593
1594 raid_end_discard_bio(r10_bio);
1595 rdev_dec_pending(rdev, conf->mddev);
1596 }
1597
1598 /*
1599 * There are some limitations to handle discard bio
1600 * 1st, the discard size is bigger than stripe_size*2.
1601 * 2st, if the discard bio spans reshape progress, we use the old way to
1602 * handle discard bio
1603 */
raid10_handle_discard(struct mddev * mddev,struct bio * bio)1604 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1605 {
1606 struct r10conf *conf = mddev->private;
1607 struct geom *geo = &conf->geo;
1608 int far_copies = geo->far_copies;
1609 bool first_copy = true;
1610 struct r10bio *r10_bio, *first_r10bio;
1611 struct bio *split;
1612 int disk;
1613 sector_t chunk;
1614 unsigned int stripe_size;
1615 unsigned int stripe_data_disks;
1616 sector_t split_size;
1617 sector_t bio_start, bio_end;
1618 sector_t first_stripe_index, last_stripe_index;
1619 sector_t start_disk_offset;
1620 unsigned int start_disk_index;
1621 sector_t end_disk_offset;
1622 unsigned int end_disk_index;
1623 unsigned int remainder;
1624
1625 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1626 return -EAGAIN;
1627
1628 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1629 bio_wouldblock_error(bio);
1630 return 0;
1631 }
1632 wait_barrier(conf, false);
1633
1634 /*
1635 * Check reshape again to avoid reshape happens after checking
1636 * MD_RECOVERY_RESHAPE and before wait_barrier
1637 */
1638 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1639 goto out;
1640
1641 if (geo->near_copies)
1642 stripe_data_disks = geo->raid_disks / geo->near_copies +
1643 geo->raid_disks % geo->near_copies;
1644 else
1645 stripe_data_disks = geo->raid_disks;
1646
1647 stripe_size = stripe_data_disks << geo->chunk_shift;
1648
1649 bio_start = bio->bi_iter.bi_sector;
1650 bio_end = bio_end_sector(bio);
1651
1652 /*
1653 * Maybe one discard bio is smaller than strip size or across one
1654 * stripe and discard region is larger than one stripe size. For far
1655 * offset layout, if the discard region is not aligned with stripe
1656 * size, there is hole when we submit discard bio to member disk.
1657 * For simplicity, we only handle discard bio which discard region
1658 * is bigger than stripe_size * 2
1659 */
1660 if (bio_sectors(bio) < stripe_size*2)
1661 goto out;
1662
1663 /*
1664 * Keep bio aligned with strip size.
1665 */
1666 div_u64_rem(bio_start, stripe_size, &remainder);
1667 if (remainder) {
1668 split_size = stripe_size - remainder;
1669 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1670 bio_chain(split, bio);
1671 allow_barrier(conf);
1672 /* Resend the fist split part */
1673 submit_bio_noacct(split);
1674 wait_barrier(conf, false);
1675 }
1676 div_u64_rem(bio_end, stripe_size, &remainder);
1677 if (remainder) {
1678 split_size = bio_sectors(bio) - remainder;
1679 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1680 bio_chain(split, bio);
1681 allow_barrier(conf);
1682 /* Resend the second split part */
1683 submit_bio_noacct(bio);
1684 bio = split;
1685 wait_barrier(conf, false);
1686 }
1687
1688 bio_start = bio->bi_iter.bi_sector;
1689 bio_end = bio_end_sector(bio);
1690
1691 /*
1692 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1693 * One stripe contains the chunks from all member disk (one chunk from
1694 * one disk at the same HBA address). For layout detail, see 'man md 4'
1695 */
1696 chunk = bio_start >> geo->chunk_shift;
1697 chunk *= geo->near_copies;
1698 first_stripe_index = chunk;
1699 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1700 if (geo->far_offset)
1701 first_stripe_index *= geo->far_copies;
1702 start_disk_offset = (bio_start & geo->chunk_mask) +
1703 (first_stripe_index << geo->chunk_shift);
1704
1705 chunk = bio_end >> geo->chunk_shift;
1706 chunk *= geo->near_copies;
1707 last_stripe_index = chunk;
1708 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1709 if (geo->far_offset)
1710 last_stripe_index *= geo->far_copies;
1711 end_disk_offset = (bio_end & geo->chunk_mask) +
1712 (last_stripe_index << geo->chunk_shift);
1713
1714 retry_discard:
1715 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1716 r10_bio->mddev = mddev;
1717 r10_bio->state = 0;
1718 r10_bio->sectors = 0;
1719 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1720 wait_blocked_dev(mddev, r10_bio);
1721
1722 /*
1723 * For far layout it needs more than one r10bio to cover all regions.
1724 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1725 * to record the discard bio. Other r10bio->master_bio record the first
1726 * r10bio. The first r10bio only release after all other r10bios finish.
1727 * The discard bio returns only first r10bio finishes
1728 */
1729 if (first_copy) {
1730 r10_bio->master_bio = bio;
1731 set_bit(R10BIO_Discard, &r10_bio->state);
1732 first_copy = false;
1733 first_r10bio = r10_bio;
1734 } else
1735 r10_bio->master_bio = (struct bio *)first_r10bio;
1736
1737 /*
1738 * first select target devices under rcu_lock and
1739 * inc refcount on their rdev. Record them by setting
1740 * bios[x] to bio
1741 */
1742 rcu_read_lock();
1743 for (disk = 0; disk < geo->raid_disks; disk++) {
1744 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1745 struct md_rdev *rrdev = rcu_dereference(
1746 conf->mirrors[disk].replacement);
1747
1748 r10_bio->devs[disk].bio = NULL;
1749 r10_bio->devs[disk].repl_bio = NULL;
1750
1751 if (rdev && (test_bit(Faulty, &rdev->flags)))
1752 rdev = NULL;
1753 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1754 rrdev = NULL;
1755 if (!rdev && !rrdev)
1756 continue;
1757
1758 if (rdev) {
1759 r10_bio->devs[disk].bio = bio;
1760 atomic_inc(&rdev->nr_pending);
1761 }
1762 if (rrdev) {
1763 r10_bio->devs[disk].repl_bio = bio;
1764 atomic_inc(&rrdev->nr_pending);
1765 }
1766 }
1767 rcu_read_unlock();
1768
1769 atomic_set(&r10_bio->remaining, 1);
1770 for (disk = 0; disk < geo->raid_disks; disk++) {
1771 sector_t dev_start, dev_end;
1772 struct bio *mbio, *rbio = NULL;
1773
1774 /*
1775 * Now start to calculate the start and end address for each disk.
1776 * The space between dev_start and dev_end is the discard region.
1777 *
1778 * For dev_start, it needs to consider three conditions:
1779 * 1st, the disk is before start_disk, you can imagine the disk in
1780 * the next stripe. So the dev_start is the start address of next
1781 * stripe.
1782 * 2st, the disk is after start_disk, it means the disk is at the
1783 * same stripe of first disk
1784 * 3st, the first disk itself, we can use start_disk_offset directly
1785 */
1786 if (disk < start_disk_index)
1787 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1788 else if (disk > start_disk_index)
1789 dev_start = first_stripe_index * mddev->chunk_sectors;
1790 else
1791 dev_start = start_disk_offset;
1792
1793 if (disk < end_disk_index)
1794 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1795 else if (disk > end_disk_index)
1796 dev_end = last_stripe_index * mddev->chunk_sectors;
1797 else
1798 dev_end = end_disk_offset;
1799
1800 /*
1801 * It only handles discard bio which size is >= stripe size, so
1802 * dev_end > dev_start all the time.
1803 * It doesn't need to use rcu lock to get rdev here. We already
1804 * add rdev->nr_pending in the first loop.
1805 */
1806 if (r10_bio->devs[disk].bio) {
1807 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1808 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1809 &mddev->bio_set);
1810 mbio->bi_end_io = raid10_end_discard_request;
1811 mbio->bi_private = r10_bio;
1812 r10_bio->devs[disk].bio = mbio;
1813 r10_bio->devs[disk].devnum = disk;
1814 atomic_inc(&r10_bio->remaining);
1815 md_submit_discard_bio(mddev, rdev, mbio,
1816 dev_start + choose_data_offset(r10_bio, rdev),
1817 dev_end - dev_start);
1818 bio_endio(mbio);
1819 }
1820 if (r10_bio->devs[disk].repl_bio) {
1821 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1822 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1823 &mddev->bio_set);
1824 rbio->bi_end_io = raid10_end_discard_request;
1825 rbio->bi_private = r10_bio;
1826 r10_bio->devs[disk].repl_bio = rbio;
1827 r10_bio->devs[disk].devnum = disk;
1828 atomic_inc(&r10_bio->remaining);
1829 md_submit_discard_bio(mddev, rrdev, rbio,
1830 dev_start + choose_data_offset(r10_bio, rrdev),
1831 dev_end - dev_start);
1832 bio_endio(rbio);
1833 }
1834 }
1835
1836 if (!geo->far_offset && --far_copies) {
1837 first_stripe_index += geo->stride >> geo->chunk_shift;
1838 start_disk_offset += geo->stride;
1839 last_stripe_index += geo->stride >> geo->chunk_shift;
1840 end_disk_offset += geo->stride;
1841 atomic_inc(&first_r10bio->remaining);
1842 raid_end_discard_bio(r10_bio);
1843 wait_barrier(conf, false);
1844 goto retry_discard;
1845 }
1846
1847 raid_end_discard_bio(r10_bio);
1848
1849 return 0;
1850 out:
1851 allow_barrier(conf);
1852 return -EAGAIN;
1853 }
1854
raid10_make_request(struct mddev * mddev,struct bio * bio)1855 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1856 {
1857 struct r10conf *conf = mddev->private;
1858 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1859 int chunk_sects = chunk_mask + 1;
1860 int sectors = bio_sectors(bio);
1861
1862 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1863 && md_flush_request(mddev, bio))
1864 return true;
1865
1866 if (!md_write_start(mddev, bio))
1867 return false;
1868
1869 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1870 if (!raid10_handle_discard(mddev, bio))
1871 return true;
1872
1873 /*
1874 * If this request crosses a chunk boundary, we need to split
1875 * it.
1876 */
1877 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1878 sectors > chunk_sects
1879 && (conf->geo.near_copies < conf->geo.raid_disks
1880 || conf->prev.near_copies <
1881 conf->prev.raid_disks)))
1882 sectors = chunk_sects -
1883 (bio->bi_iter.bi_sector &
1884 (chunk_sects - 1));
1885 __make_request(mddev, bio, sectors);
1886
1887 /* In case raid10d snuck in to freeze_array */
1888 wake_up(&conf->wait_barrier);
1889 return true;
1890 }
1891
raid10_status(struct seq_file * seq,struct mddev * mddev)1892 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1893 {
1894 struct r10conf *conf = mddev->private;
1895 int i;
1896
1897 if (conf->geo.near_copies < conf->geo.raid_disks)
1898 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1899 if (conf->geo.near_copies > 1)
1900 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1901 if (conf->geo.far_copies > 1) {
1902 if (conf->geo.far_offset)
1903 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1904 else
1905 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1906 if (conf->geo.far_set_size != conf->geo.raid_disks)
1907 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1908 }
1909 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1910 conf->geo.raid_disks - mddev->degraded);
1911 rcu_read_lock();
1912 for (i = 0; i < conf->geo.raid_disks; i++) {
1913 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1914 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1915 }
1916 rcu_read_unlock();
1917 seq_printf(seq, "]");
1918 }
1919
1920 /* check if there are enough drives for
1921 * every block to appear on atleast one.
1922 * Don't consider the device numbered 'ignore'
1923 * as we might be about to remove it.
1924 */
_enough(struct r10conf * conf,int previous,int ignore)1925 static int _enough(struct r10conf *conf, int previous, int ignore)
1926 {
1927 int first = 0;
1928 int has_enough = 0;
1929 int disks, ncopies;
1930 if (previous) {
1931 disks = conf->prev.raid_disks;
1932 ncopies = conf->prev.near_copies;
1933 } else {
1934 disks = conf->geo.raid_disks;
1935 ncopies = conf->geo.near_copies;
1936 }
1937
1938 rcu_read_lock();
1939 do {
1940 int n = conf->copies;
1941 int cnt = 0;
1942 int this = first;
1943 while (n--) {
1944 struct md_rdev *rdev;
1945 if (this != ignore &&
1946 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1947 test_bit(In_sync, &rdev->flags))
1948 cnt++;
1949 this = (this+1) % disks;
1950 }
1951 if (cnt == 0)
1952 goto out;
1953 first = (first + ncopies) % disks;
1954 } while (first != 0);
1955 has_enough = 1;
1956 out:
1957 rcu_read_unlock();
1958 return has_enough;
1959 }
1960
enough(struct r10conf * conf,int ignore)1961 static int enough(struct r10conf *conf, int ignore)
1962 {
1963 /* when calling 'enough', both 'prev' and 'geo' must
1964 * be stable.
1965 * This is ensured if ->reconfig_mutex or ->device_lock
1966 * is held.
1967 */
1968 return _enough(conf, 0, ignore) &&
1969 _enough(conf, 1, ignore);
1970 }
1971
1972 /**
1973 * raid10_error() - RAID10 error handler.
1974 * @mddev: affected md device.
1975 * @rdev: member device to fail.
1976 *
1977 * The routine acknowledges &rdev failure and determines new @mddev state.
1978 * If it failed, then:
1979 * - &MD_BROKEN flag is set in &mddev->flags.
1980 * Otherwise, it must be degraded:
1981 * - recovery is interrupted.
1982 * - &mddev->degraded is bumped.
1983
1984 * @rdev is marked as &Faulty excluding case when array is failed and
1985 * &mddev->fail_last_dev is off.
1986 */
raid10_error(struct mddev * mddev,struct md_rdev * rdev)1987 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1988 {
1989 struct r10conf *conf = mddev->private;
1990 unsigned long flags;
1991
1992 spin_lock_irqsave(&conf->device_lock, flags);
1993
1994 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
1995 set_bit(MD_BROKEN, &mddev->flags);
1996
1997 if (!mddev->fail_last_dev) {
1998 spin_unlock_irqrestore(&conf->device_lock, flags);
1999 return;
2000 }
2001 }
2002 if (test_and_clear_bit(In_sync, &rdev->flags))
2003 mddev->degraded++;
2004
2005 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2006 set_bit(Blocked, &rdev->flags);
2007 set_bit(Faulty, &rdev->flags);
2008 set_mask_bits(&mddev->sb_flags, 0,
2009 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2010 spin_unlock_irqrestore(&conf->device_lock, flags);
2011 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2012 "md/raid10:%s: Operation continuing on %d devices.\n",
2013 mdname(mddev), rdev->bdev,
2014 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2015 }
2016
print_conf(struct r10conf * conf)2017 static void print_conf(struct r10conf *conf)
2018 {
2019 int i;
2020 struct md_rdev *rdev;
2021
2022 pr_debug("RAID10 conf printout:\n");
2023 if (!conf) {
2024 pr_debug("(!conf)\n");
2025 return;
2026 }
2027 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2028 conf->geo.raid_disks);
2029
2030 /* This is only called with ->reconfix_mutex held, so
2031 * rcu protection of rdev is not needed */
2032 for (i = 0; i < conf->geo.raid_disks; i++) {
2033 rdev = conf->mirrors[i].rdev;
2034 if (rdev)
2035 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2036 i, !test_bit(In_sync, &rdev->flags),
2037 !test_bit(Faulty, &rdev->flags),
2038 rdev->bdev);
2039 }
2040 }
2041
close_sync(struct r10conf * conf)2042 static void close_sync(struct r10conf *conf)
2043 {
2044 wait_barrier(conf, false);
2045 allow_barrier(conf);
2046
2047 mempool_exit(&conf->r10buf_pool);
2048 }
2049
raid10_spare_active(struct mddev * mddev)2050 static int raid10_spare_active(struct mddev *mddev)
2051 {
2052 int i;
2053 struct r10conf *conf = mddev->private;
2054 struct raid10_info *tmp;
2055 int count = 0;
2056 unsigned long flags;
2057
2058 /*
2059 * Find all non-in_sync disks within the RAID10 configuration
2060 * and mark them in_sync
2061 */
2062 for (i = 0; i < conf->geo.raid_disks; i++) {
2063 tmp = conf->mirrors + i;
2064 if (tmp->replacement
2065 && tmp->replacement->recovery_offset == MaxSector
2066 && !test_bit(Faulty, &tmp->replacement->flags)
2067 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2068 /* Replacement has just become active */
2069 if (!tmp->rdev
2070 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2071 count++;
2072 if (tmp->rdev) {
2073 /* Replaced device not technically faulty,
2074 * but we need to be sure it gets removed
2075 * and never re-added.
2076 */
2077 set_bit(Faulty, &tmp->rdev->flags);
2078 sysfs_notify_dirent_safe(
2079 tmp->rdev->sysfs_state);
2080 }
2081 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2082 } else if (tmp->rdev
2083 && tmp->rdev->recovery_offset == MaxSector
2084 && !test_bit(Faulty, &tmp->rdev->flags)
2085 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2086 count++;
2087 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2088 }
2089 }
2090 spin_lock_irqsave(&conf->device_lock, flags);
2091 mddev->degraded -= count;
2092 spin_unlock_irqrestore(&conf->device_lock, flags);
2093
2094 print_conf(conf);
2095 return count;
2096 }
2097
raid10_add_disk(struct mddev * mddev,struct md_rdev * rdev)2098 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2099 {
2100 struct r10conf *conf = mddev->private;
2101 int err = -EEXIST;
2102 int mirror;
2103 int first = 0;
2104 int last = conf->geo.raid_disks - 1;
2105
2106 if (mddev->recovery_cp < MaxSector)
2107 /* only hot-add to in-sync arrays, as recovery is
2108 * very different from resync
2109 */
2110 return -EBUSY;
2111 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2112 return -EINVAL;
2113
2114 if (md_integrity_add_rdev(rdev, mddev))
2115 return -ENXIO;
2116
2117 if (rdev->raid_disk >= 0)
2118 first = last = rdev->raid_disk;
2119
2120 if (rdev->saved_raid_disk >= first &&
2121 rdev->saved_raid_disk < conf->geo.raid_disks &&
2122 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2123 mirror = rdev->saved_raid_disk;
2124 else
2125 mirror = first;
2126 for ( ; mirror <= last ; mirror++) {
2127 struct raid10_info *p = &conf->mirrors[mirror];
2128 if (p->recovery_disabled == mddev->recovery_disabled)
2129 continue;
2130 if (p->rdev) {
2131 if (!test_bit(WantReplacement, &p->rdev->flags) ||
2132 p->replacement != NULL)
2133 continue;
2134 clear_bit(In_sync, &rdev->flags);
2135 set_bit(Replacement, &rdev->flags);
2136 rdev->raid_disk = mirror;
2137 err = 0;
2138 if (mddev->gendisk)
2139 disk_stack_limits(mddev->gendisk, rdev->bdev,
2140 rdev->data_offset << 9);
2141 conf->fullsync = 1;
2142 rcu_assign_pointer(p->replacement, rdev);
2143 break;
2144 }
2145
2146 if (mddev->gendisk)
2147 disk_stack_limits(mddev->gendisk, rdev->bdev,
2148 rdev->data_offset << 9);
2149
2150 p->head_position = 0;
2151 p->recovery_disabled = mddev->recovery_disabled - 1;
2152 rdev->raid_disk = mirror;
2153 err = 0;
2154 if (rdev->saved_raid_disk != mirror)
2155 conf->fullsync = 1;
2156 rcu_assign_pointer(p->rdev, rdev);
2157 break;
2158 }
2159
2160 print_conf(conf);
2161 return err;
2162 }
2163
raid10_remove_disk(struct mddev * mddev,struct md_rdev * rdev)2164 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2165 {
2166 struct r10conf *conf = mddev->private;
2167 int err = 0;
2168 int number = rdev->raid_disk;
2169 struct md_rdev **rdevp;
2170 struct raid10_info *p;
2171
2172 print_conf(conf);
2173 if (unlikely(number >= mddev->raid_disks))
2174 return 0;
2175 p = conf->mirrors + number;
2176 if (rdev == p->rdev)
2177 rdevp = &p->rdev;
2178 else if (rdev == p->replacement)
2179 rdevp = &p->replacement;
2180 else
2181 return 0;
2182
2183 if (test_bit(In_sync, &rdev->flags) ||
2184 atomic_read(&rdev->nr_pending)) {
2185 err = -EBUSY;
2186 goto abort;
2187 }
2188 /* Only remove non-faulty devices if recovery
2189 * is not possible.
2190 */
2191 if (!test_bit(Faulty, &rdev->flags) &&
2192 mddev->recovery_disabled != p->recovery_disabled &&
2193 (!p->replacement || p->replacement == rdev) &&
2194 number < conf->geo.raid_disks &&
2195 enough(conf, -1)) {
2196 err = -EBUSY;
2197 goto abort;
2198 }
2199 *rdevp = NULL;
2200 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
2201 synchronize_rcu();
2202 if (atomic_read(&rdev->nr_pending)) {
2203 /* lost the race, try later */
2204 err = -EBUSY;
2205 *rdevp = rdev;
2206 goto abort;
2207 }
2208 }
2209 if (p->replacement) {
2210 /* We must have just cleared 'rdev' */
2211 p->rdev = p->replacement;
2212 clear_bit(Replacement, &p->replacement->flags);
2213 smp_mb(); /* Make sure other CPUs may see both as identical
2214 * but will never see neither -- if they are careful.
2215 */
2216 p->replacement = NULL;
2217 }
2218
2219 clear_bit(WantReplacement, &rdev->flags);
2220 err = md_integrity_register(mddev);
2221
2222 abort:
2223
2224 print_conf(conf);
2225 return err;
2226 }
2227
__end_sync_read(struct r10bio * r10_bio,struct bio * bio,int d)2228 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2229 {
2230 struct r10conf *conf = r10_bio->mddev->private;
2231
2232 if (!bio->bi_status)
2233 set_bit(R10BIO_Uptodate, &r10_bio->state);
2234 else
2235 /* The write handler will notice the lack of
2236 * R10BIO_Uptodate and record any errors etc
2237 */
2238 atomic_add(r10_bio->sectors,
2239 &conf->mirrors[d].rdev->corrected_errors);
2240
2241 /* for reconstruct, we always reschedule after a read.
2242 * for resync, only after all reads
2243 */
2244 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2245 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2246 atomic_dec_and_test(&r10_bio->remaining)) {
2247 /* we have read all the blocks,
2248 * do the comparison in process context in raid10d
2249 */
2250 reschedule_retry(r10_bio);
2251 }
2252 }
2253
end_sync_read(struct bio * bio)2254 static void end_sync_read(struct bio *bio)
2255 {
2256 struct r10bio *r10_bio = get_resync_r10bio(bio);
2257 struct r10conf *conf = r10_bio->mddev->private;
2258 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2259
2260 __end_sync_read(r10_bio, bio, d);
2261 }
2262
end_reshape_read(struct bio * bio)2263 static void end_reshape_read(struct bio *bio)
2264 {
2265 /* reshape read bio isn't allocated from r10buf_pool */
2266 struct r10bio *r10_bio = bio->bi_private;
2267
2268 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2269 }
2270
end_sync_request(struct r10bio * r10_bio)2271 static void end_sync_request(struct r10bio *r10_bio)
2272 {
2273 struct mddev *mddev = r10_bio->mddev;
2274
2275 while (atomic_dec_and_test(&r10_bio->remaining)) {
2276 if (r10_bio->master_bio == NULL) {
2277 /* the primary of several recovery bios */
2278 sector_t s = r10_bio->sectors;
2279 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2280 test_bit(R10BIO_WriteError, &r10_bio->state))
2281 reschedule_retry(r10_bio);
2282 else
2283 put_buf(r10_bio);
2284 md_done_sync(mddev, s, 1);
2285 break;
2286 } else {
2287 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2288 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2289 test_bit(R10BIO_WriteError, &r10_bio->state))
2290 reschedule_retry(r10_bio);
2291 else
2292 put_buf(r10_bio);
2293 r10_bio = r10_bio2;
2294 }
2295 }
2296 }
2297
end_sync_write(struct bio * bio)2298 static void end_sync_write(struct bio *bio)
2299 {
2300 struct r10bio *r10_bio = get_resync_r10bio(bio);
2301 struct mddev *mddev = r10_bio->mddev;
2302 struct r10conf *conf = mddev->private;
2303 int d;
2304 sector_t first_bad;
2305 int bad_sectors;
2306 int slot;
2307 int repl;
2308 struct md_rdev *rdev = NULL;
2309
2310 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2311 if (repl)
2312 rdev = conf->mirrors[d].replacement;
2313 else
2314 rdev = conf->mirrors[d].rdev;
2315
2316 if (bio->bi_status) {
2317 if (repl)
2318 md_error(mddev, rdev);
2319 else {
2320 set_bit(WriteErrorSeen, &rdev->flags);
2321 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2322 set_bit(MD_RECOVERY_NEEDED,
2323 &rdev->mddev->recovery);
2324 set_bit(R10BIO_WriteError, &r10_bio->state);
2325 }
2326 } else if (is_badblock(rdev,
2327 r10_bio->devs[slot].addr,
2328 r10_bio->sectors,
2329 &first_bad, &bad_sectors))
2330 set_bit(R10BIO_MadeGood, &r10_bio->state);
2331
2332 rdev_dec_pending(rdev, mddev);
2333
2334 end_sync_request(r10_bio);
2335 }
2336
2337 /*
2338 * Note: sync and recover and handled very differently for raid10
2339 * This code is for resync.
2340 * For resync, we read through virtual addresses and read all blocks.
2341 * If there is any error, we schedule a write. The lowest numbered
2342 * drive is authoritative.
2343 * However requests come for physical address, so we need to map.
2344 * For every physical address there are raid_disks/copies virtual addresses,
2345 * which is always are least one, but is not necessarly an integer.
2346 * This means that a physical address can span multiple chunks, so we may
2347 * have to submit multiple io requests for a single sync request.
2348 */
2349 /*
2350 * We check if all blocks are in-sync and only write to blocks that
2351 * aren't in sync
2352 */
sync_request_write(struct mddev * mddev,struct r10bio * r10_bio)2353 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2354 {
2355 struct r10conf *conf = mddev->private;
2356 int i, first;
2357 struct bio *tbio, *fbio;
2358 int vcnt;
2359 struct page **tpages, **fpages;
2360
2361 atomic_set(&r10_bio->remaining, 1);
2362
2363 /* find the first device with a block */
2364 for (i=0; i<conf->copies; i++)
2365 if (!r10_bio->devs[i].bio->bi_status)
2366 break;
2367
2368 if (i == conf->copies)
2369 goto done;
2370
2371 first = i;
2372 fbio = r10_bio->devs[i].bio;
2373 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2374 fbio->bi_iter.bi_idx = 0;
2375 fpages = get_resync_pages(fbio)->pages;
2376
2377 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2378 /* now find blocks with errors */
2379 for (i=0 ; i < conf->copies ; i++) {
2380 int j, d;
2381 struct md_rdev *rdev;
2382 struct resync_pages *rp;
2383
2384 tbio = r10_bio->devs[i].bio;
2385
2386 if (tbio->bi_end_io != end_sync_read)
2387 continue;
2388 if (i == first)
2389 continue;
2390
2391 tpages = get_resync_pages(tbio)->pages;
2392 d = r10_bio->devs[i].devnum;
2393 rdev = conf->mirrors[d].rdev;
2394 if (!r10_bio->devs[i].bio->bi_status) {
2395 /* We know that the bi_io_vec layout is the same for
2396 * both 'first' and 'i', so we just compare them.
2397 * All vec entries are PAGE_SIZE;
2398 */
2399 int sectors = r10_bio->sectors;
2400 for (j = 0; j < vcnt; j++) {
2401 int len = PAGE_SIZE;
2402 if (sectors < (len / 512))
2403 len = sectors * 512;
2404 if (memcmp(page_address(fpages[j]),
2405 page_address(tpages[j]),
2406 len))
2407 break;
2408 sectors -= len/512;
2409 }
2410 if (j == vcnt)
2411 continue;
2412 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2413 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2414 /* Don't fix anything. */
2415 continue;
2416 } else if (test_bit(FailFast, &rdev->flags)) {
2417 /* Just give up on this device */
2418 md_error(rdev->mddev, rdev);
2419 continue;
2420 }
2421 /* Ok, we need to write this bio, either to correct an
2422 * inconsistency or to correct an unreadable block.
2423 * First we need to fixup bv_offset, bv_len and
2424 * bi_vecs, as the read request might have corrupted these
2425 */
2426 rp = get_resync_pages(tbio);
2427 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2428
2429 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2430
2431 rp->raid_bio = r10_bio;
2432 tbio->bi_private = rp;
2433 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2434 tbio->bi_end_io = end_sync_write;
2435
2436 bio_copy_data(tbio, fbio);
2437
2438 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2439 atomic_inc(&r10_bio->remaining);
2440 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2441
2442 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2443 tbio->bi_opf |= MD_FAILFAST;
2444 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2445 submit_bio_noacct(tbio);
2446 }
2447
2448 /* Now write out to any replacement devices
2449 * that are active
2450 */
2451 for (i = 0; i < conf->copies; i++) {
2452 int d;
2453
2454 tbio = r10_bio->devs[i].repl_bio;
2455 if (!tbio || !tbio->bi_end_io)
2456 continue;
2457 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2458 && r10_bio->devs[i].bio != fbio)
2459 bio_copy_data(tbio, fbio);
2460 d = r10_bio->devs[i].devnum;
2461 atomic_inc(&r10_bio->remaining);
2462 md_sync_acct(conf->mirrors[d].replacement->bdev,
2463 bio_sectors(tbio));
2464 submit_bio_noacct(tbio);
2465 }
2466
2467 done:
2468 if (atomic_dec_and_test(&r10_bio->remaining)) {
2469 md_done_sync(mddev, r10_bio->sectors, 1);
2470 put_buf(r10_bio);
2471 }
2472 }
2473
2474 /*
2475 * Now for the recovery code.
2476 * Recovery happens across physical sectors.
2477 * We recover all non-is_sync drives by finding the virtual address of
2478 * each, and then choose a working drive that also has that virt address.
2479 * There is a separate r10_bio for each non-in_sync drive.
2480 * Only the first two slots are in use. The first for reading,
2481 * The second for writing.
2482 *
2483 */
fix_recovery_read_error(struct r10bio * r10_bio)2484 static void fix_recovery_read_error(struct r10bio *r10_bio)
2485 {
2486 /* We got a read error during recovery.
2487 * We repeat the read in smaller page-sized sections.
2488 * If a read succeeds, write it to the new device or record
2489 * a bad block if we cannot.
2490 * If a read fails, record a bad block on both old and
2491 * new devices.
2492 */
2493 struct mddev *mddev = r10_bio->mddev;
2494 struct r10conf *conf = mddev->private;
2495 struct bio *bio = r10_bio->devs[0].bio;
2496 sector_t sect = 0;
2497 int sectors = r10_bio->sectors;
2498 int idx = 0;
2499 int dr = r10_bio->devs[0].devnum;
2500 int dw = r10_bio->devs[1].devnum;
2501 struct page **pages = get_resync_pages(bio)->pages;
2502
2503 while (sectors) {
2504 int s = sectors;
2505 struct md_rdev *rdev;
2506 sector_t addr;
2507 int ok;
2508
2509 if (s > (PAGE_SIZE>>9))
2510 s = PAGE_SIZE >> 9;
2511
2512 rdev = conf->mirrors[dr].rdev;
2513 addr = r10_bio->devs[0].addr + sect,
2514 ok = sync_page_io(rdev,
2515 addr,
2516 s << 9,
2517 pages[idx],
2518 REQ_OP_READ, 0, false);
2519 if (ok) {
2520 rdev = conf->mirrors[dw].rdev;
2521 addr = r10_bio->devs[1].addr + sect;
2522 ok = sync_page_io(rdev,
2523 addr,
2524 s << 9,
2525 pages[idx],
2526 REQ_OP_WRITE, 0, false);
2527 if (!ok) {
2528 set_bit(WriteErrorSeen, &rdev->flags);
2529 if (!test_and_set_bit(WantReplacement,
2530 &rdev->flags))
2531 set_bit(MD_RECOVERY_NEEDED,
2532 &rdev->mddev->recovery);
2533 }
2534 }
2535 if (!ok) {
2536 /* We don't worry if we cannot set a bad block -
2537 * it really is bad so there is no loss in not
2538 * recording it yet
2539 */
2540 rdev_set_badblocks(rdev, addr, s, 0);
2541
2542 if (rdev != conf->mirrors[dw].rdev) {
2543 /* need bad block on destination too */
2544 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2545 addr = r10_bio->devs[1].addr + sect;
2546 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2547 if (!ok) {
2548 /* just abort the recovery */
2549 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2550 mdname(mddev));
2551
2552 conf->mirrors[dw].recovery_disabled
2553 = mddev->recovery_disabled;
2554 set_bit(MD_RECOVERY_INTR,
2555 &mddev->recovery);
2556 break;
2557 }
2558 }
2559 }
2560
2561 sectors -= s;
2562 sect += s;
2563 idx++;
2564 }
2565 }
2566
recovery_request_write(struct mddev * mddev,struct r10bio * r10_bio)2567 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2568 {
2569 struct r10conf *conf = mddev->private;
2570 int d;
2571 struct bio *wbio, *wbio2;
2572
2573 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2574 fix_recovery_read_error(r10_bio);
2575 end_sync_request(r10_bio);
2576 return;
2577 }
2578
2579 /*
2580 * share the pages with the first bio
2581 * and submit the write request
2582 */
2583 d = r10_bio->devs[1].devnum;
2584 wbio = r10_bio->devs[1].bio;
2585 wbio2 = r10_bio->devs[1].repl_bio;
2586 /* Need to test wbio2->bi_end_io before we call
2587 * submit_bio_noacct as if the former is NULL,
2588 * the latter is free to free wbio2.
2589 */
2590 if (wbio2 && !wbio2->bi_end_io)
2591 wbio2 = NULL;
2592 if (wbio->bi_end_io) {
2593 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2594 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2595 submit_bio_noacct(wbio);
2596 }
2597 if (wbio2) {
2598 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2599 md_sync_acct(conf->mirrors[d].replacement->bdev,
2600 bio_sectors(wbio2));
2601 submit_bio_noacct(wbio2);
2602 }
2603 }
2604
2605 /*
2606 * Used by fix_read_error() to decay the per rdev read_errors.
2607 * We halve the read error count for every hour that has elapsed
2608 * since the last recorded read error.
2609 *
2610 */
check_decay_read_errors(struct mddev * mddev,struct md_rdev * rdev)2611 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2612 {
2613 long cur_time_mon;
2614 unsigned long hours_since_last;
2615 unsigned int read_errors = atomic_read(&rdev->read_errors);
2616
2617 cur_time_mon = ktime_get_seconds();
2618
2619 if (rdev->last_read_error == 0) {
2620 /* first time we've seen a read error */
2621 rdev->last_read_error = cur_time_mon;
2622 return;
2623 }
2624
2625 hours_since_last = (long)(cur_time_mon -
2626 rdev->last_read_error) / 3600;
2627
2628 rdev->last_read_error = cur_time_mon;
2629
2630 /*
2631 * if hours_since_last is > the number of bits in read_errors
2632 * just set read errors to 0. We do this to avoid
2633 * overflowing the shift of read_errors by hours_since_last.
2634 */
2635 if (hours_since_last >= 8 * sizeof(read_errors))
2636 atomic_set(&rdev->read_errors, 0);
2637 else
2638 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2639 }
2640
r10_sync_page_io(struct md_rdev * rdev,sector_t sector,int sectors,struct page * page,int rw)2641 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2642 int sectors, struct page *page, int rw)
2643 {
2644 sector_t first_bad;
2645 int bad_sectors;
2646
2647 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2648 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2649 return -1;
2650 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2651 /* success */
2652 return 1;
2653 if (rw == WRITE) {
2654 set_bit(WriteErrorSeen, &rdev->flags);
2655 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2656 set_bit(MD_RECOVERY_NEEDED,
2657 &rdev->mddev->recovery);
2658 }
2659 /* need to record an error - either for the block or the device */
2660 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2661 md_error(rdev->mddev, rdev);
2662 return 0;
2663 }
2664
2665 /*
2666 * This is a kernel thread which:
2667 *
2668 * 1. Retries failed read operations on working mirrors.
2669 * 2. Updates the raid superblock when problems encounter.
2670 * 3. Performs writes following reads for array synchronising.
2671 */
2672
fix_read_error(struct r10conf * conf,struct mddev * mddev,struct r10bio * r10_bio)2673 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2674 {
2675 int sect = 0; /* Offset from r10_bio->sector */
2676 int sectors = r10_bio->sectors;
2677 struct md_rdev *rdev;
2678 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2679 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2680
2681 /* still own a reference to this rdev, so it cannot
2682 * have been cleared recently.
2683 */
2684 rdev = conf->mirrors[d].rdev;
2685
2686 if (test_bit(Faulty, &rdev->flags))
2687 /* drive has already been failed, just ignore any
2688 more fix_read_error() attempts */
2689 return;
2690
2691 check_decay_read_errors(mddev, rdev);
2692 atomic_inc(&rdev->read_errors);
2693 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2694 pr_notice("md/raid10:%s: %pg: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2695 mdname(mddev), rdev->bdev,
2696 atomic_read(&rdev->read_errors), max_read_errors);
2697 pr_notice("md/raid10:%s: %pg: Failing raid device\n",
2698 mdname(mddev), rdev->bdev);
2699 md_error(mddev, rdev);
2700 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2701 return;
2702 }
2703
2704 while(sectors) {
2705 int s = sectors;
2706 int sl = r10_bio->read_slot;
2707 int success = 0;
2708 int start;
2709
2710 if (s > (PAGE_SIZE>>9))
2711 s = PAGE_SIZE >> 9;
2712
2713 rcu_read_lock();
2714 do {
2715 sector_t first_bad;
2716 int bad_sectors;
2717
2718 d = r10_bio->devs[sl].devnum;
2719 rdev = rcu_dereference(conf->mirrors[d].rdev);
2720 if (rdev &&
2721 test_bit(In_sync, &rdev->flags) &&
2722 !test_bit(Faulty, &rdev->flags) &&
2723 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2724 &first_bad, &bad_sectors) == 0) {
2725 atomic_inc(&rdev->nr_pending);
2726 rcu_read_unlock();
2727 success = sync_page_io(rdev,
2728 r10_bio->devs[sl].addr +
2729 sect,
2730 s<<9,
2731 conf->tmppage,
2732 REQ_OP_READ, 0, false);
2733 rdev_dec_pending(rdev, mddev);
2734 rcu_read_lock();
2735 if (success)
2736 break;
2737 }
2738 sl++;
2739 if (sl == conf->copies)
2740 sl = 0;
2741 } while (!success && sl != r10_bio->read_slot);
2742 rcu_read_unlock();
2743
2744 if (!success) {
2745 /* Cannot read from anywhere, just mark the block
2746 * as bad on the first device to discourage future
2747 * reads.
2748 */
2749 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2750 rdev = conf->mirrors[dn].rdev;
2751
2752 if (!rdev_set_badblocks(
2753 rdev,
2754 r10_bio->devs[r10_bio->read_slot].addr
2755 + sect,
2756 s, 0)) {
2757 md_error(mddev, rdev);
2758 r10_bio->devs[r10_bio->read_slot].bio
2759 = IO_BLOCKED;
2760 }
2761 break;
2762 }
2763
2764 start = sl;
2765 /* write it back and re-read */
2766 rcu_read_lock();
2767 while (sl != r10_bio->read_slot) {
2768 if (sl==0)
2769 sl = conf->copies;
2770 sl--;
2771 d = r10_bio->devs[sl].devnum;
2772 rdev = rcu_dereference(conf->mirrors[d].rdev);
2773 if (!rdev ||
2774 test_bit(Faulty, &rdev->flags) ||
2775 !test_bit(In_sync, &rdev->flags))
2776 continue;
2777
2778 atomic_inc(&rdev->nr_pending);
2779 rcu_read_unlock();
2780 if (r10_sync_page_io(rdev,
2781 r10_bio->devs[sl].addr +
2782 sect,
2783 s, conf->tmppage, WRITE)
2784 == 0) {
2785 /* Well, this device is dead */
2786 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2787 mdname(mddev), s,
2788 (unsigned long long)(
2789 sect +
2790 choose_data_offset(r10_bio,
2791 rdev)),
2792 rdev->bdev);
2793 pr_notice("md/raid10:%s: %pg: failing drive\n",
2794 mdname(mddev),
2795 rdev->bdev);
2796 }
2797 rdev_dec_pending(rdev, mddev);
2798 rcu_read_lock();
2799 }
2800 sl = start;
2801 while (sl != r10_bio->read_slot) {
2802 if (sl==0)
2803 sl = conf->copies;
2804 sl--;
2805 d = r10_bio->devs[sl].devnum;
2806 rdev = rcu_dereference(conf->mirrors[d].rdev);
2807 if (!rdev ||
2808 test_bit(Faulty, &rdev->flags) ||
2809 !test_bit(In_sync, &rdev->flags))
2810 continue;
2811
2812 atomic_inc(&rdev->nr_pending);
2813 rcu_read_unlock();
2814 switch (r10_sync_page_io(rdev,
2815 r10_bio->devs[sl].addr +
2816 sect,
2817 s, conf->tmppage,
2818 READ)) {
2819 case 0:
2820 /* Well, this device is dead */
2821 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2822 mdname(mddev), s,
2823 (unsigned long long)(
2824 sect +
2825 choose_data_offset(r10_bio, rdev)),
2826 rdev->bdev);
2827 pr_notice("md/raid10:%s: %pg: failing drive\n",
2828 mdname(mddev),
2829 rdev->bdev);
2830 break;
2831 case 1:
2832 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2833 mdname(mddev), s,
2834 (unsigned long long)(
2835 sect +
2836 choose_data_offset(r10_bio, rdev)),
2837 rdev->bdev);
2838 atomic_add(s, &rdev->corrected_errors);
2839 }
2840
2841 rdev_dec_pending(rdev, mddev);
2842 rcu_read_lock();
2843 }
2844 rcu_read_unlock();
2845
2846 sectors -= s;
2847 sect += s;
2848 }
2849 }
2850
narrow_write_error(struct r10bio * r10_bio,int i)2851 static int narrow_write_error(struct r10bio *r10_bio, int i)
2852 {
2853 struct bio *bio = r10_bio->master_bio;
2854 struct mddev *mddev = r10_bio->mddev;
2855 struct r10conf *conf = mddev->private;
2856 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2857 /* bio has the data to be written to slot 'i' where
2858 * we just recently had a write error.
2859 * We repeatedly clone the bio and trim down to one block,
2860 * then try the write. Where the write fails we record
2861 * a bad block.
2862 * It is conceivable that the bio doesn't exactly align with
2863 * blocks. We must handle this.
2864 *
2865 * We currently own a reference to the rdev.
2866 */
2867
2868 int block_sectors;
2869 sector_t sector;
2870 int sectors;
2871 int sect_to_write = r10_bio->sectors;
2872 int ok = 1;
2873
2874 if (rdev->badblocks.shift < 0)
2875 return 0;
2876
2877 block_sectors = roundup(1 << rdev->badblocks.shift,
2878 bdev_logical_block_size(rdev->bdev) >> 9);
2879 sector = r10_bio->sector;
2880 sectors = ((r10_bio->sector + block_sectors)
2881 & ~(sector_t)(block_sectors - 1))
2882 - sector;
2883
2884 while (sect_to_write) {
2885 struct bio *wbio;
2886 sector_t wsector;
2887 if (sectors > sect_to_write)
2888 sectors = sect_to_write;
2889 /* Write at 'sector' for 'sectors' */
2890 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2891 &mddev->bio_set);
2892 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2893 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2894 wbio->bi_iter.bi_sector = wsector +
2895 choose_data_offset(r10_bio, rdev);
2896 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2897
2898 if (submit_bio_wait(wbio) < 0)
2899 /* Failure! */
2900 ok = rdev_set_badblocks(rdev, wsector,
2901 sectors, 0)
2902 && ok;
2903
2904 bio_put(wbio);
2905 sect_to_write -= sectors;
2906 sector += sectors;
2907 sectors = block_sectors;
2908 }
2909 return ok;
2910 }
2911
handle_read_error(struct mddev * mddev,struct r10bio * r10_bio)2912 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2913 {
2914 int slot = r10_bio->read_slot;
2915 struct bio *bio;
2916 struct r10conf *conf = mddev->private;
2917 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2918
2919 /* we got a read error. Maybe the drive is bad. Maybe just
2920 * the block and we can fix it.
2921 * We freeze all other IO, and try reading the block from
2922 * other devices. When we find one, we re-write
2923 * and check it that fixes the read error.
2924 * This is all done synchronously while the array is
2925 * frozen.
2926 */
2927 bio = r10_bio->devs[slot].bio;
2928 bio_put(bio);
2929 r10_bio->devs[slot].bio = NULL;
2930
2931 if (mddev->ro)
2932 r10_bio->devs[slot].bio = IO_BLOCKED;
2933 else if (!test_bit(FailFast, &rdev->flags)) {
2934 freeze_array(conf, 1);
2935 fix_read_error(conf, mddev, r10_bio);
2936 unfreeze_array(conf);
2937 } else
2938 md_error(mddev, rdev);
2939
2940 rdev_dec_pending(rdev, mddev);
2941 allow_barrier(conf);
2942 r10_bio->state = 0;
2943 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2944 }
2945
handle_write_completed(struct r10conf * conf,struct r10bio * r10_bio)2946 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2947 {
2948 /* Some sort of write request has finished and it
2949 * succeeded in writing where we thought there was a
2950 * bad block. So forget the bad block.
2951 * Or possibly if failed and we need to record
2952 * a bad block.
2953 */
2954 int m;
2955 struct md_rdev *rdev;
2956
2957 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2958 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2959 for (m = 0; m < conf->copies; m++) {
2960 int dev = r10_bio->devs[m].devnum;
2961 rdev = conf->mirrors[dev].rdev;
2962 if (r10_bio->devs[m].bio == NULL ||
2963 r10_bio->devs[m].bio->bi_end_io == NULL)
2964 continue;
2965 if (!r10_bio->devs[m].bio->bi_status) {
2966 rdev_clear_badblocks(
2967 rdev,
2968 r10_bio->devs[m].addr,
2969 r10_bio->sectors, 0);
2970 } else {
2971 if (!rdev_set_badblocks(
2972 rdev,
2973 r10_bio->devs[m].addr,
2974 r10_bio->sectors, 0))
2975 md_error(conf->mddev, rdev);
2976 }
2977 rdev = conf->mirrors[dev].replacement;
2978 if (r10_bio->devs[m].repl_bio == NULL ||
2979 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2980 continue;
2981
2982 if (!r10_bio->devs[m].repl_bio->bi_status) {
2983 rdev_clear_badblocks(
2984 rdev,
2985 r10_bio->devs[m].addr,
2986 r10_bio->sectors, 0);
2987 } else {
2988 if (!rdev_set_badblocks(
2989 rdev,
2990 r10_bio->devs[m].addr,
2991 r10_bio->sectors, 0))
2992 md_error(conf->mddev, rdev);
2993 }
2994 }
2995 put_buf(r10_bio);
2996 } else {
2997 bool fail = false;
2998 for (m = 0; m < conf->copies; m++) {
2999 int dev = r10_bio->devs[m].devnum;
3000 struct bio *bio = r10_bio->devs[m].bio;
3001 rdev = conf->mirrors[dev].rdev;
3002 if (bio == IO_MADE_GOOD) {
3003 rdev_clear_badblocks(
3004 rdev,
3005 r10_bio->devs[m].addr,
3006 r10_bio->sectors, 0);
3007 rdev_dec_pending(rdev, conf->mddev);
3008 } else if (bio != NULL && bio->bi_status) {
3009 fail = true;
3010 if (!narrow_write_error(r10_bio, m)) {
3011 md_error(conf->mddev, rdev);
3012 set_bit(R10BIO_Degraded,
3013 &r10_bio->state);
3014 }
3015 rdev_dec_pending(rdev, conf->mddev);
3016 }
3017 bio = r10_bio->devs[m].repl_bio;
3018 rdev = conf->mirrors[dev].replacement;
3019 if (rdev && bio == IO_MADE_GOOD) {
3020 rdev_clear_badblocks(
3021 rdev,
3022 r10_bio->devs[m].addr,
3023 r10_bio->sectors, 0);
3024 rdev_dec_pending(rdev, conf->mddev);
3025 }
3026 }
3027 if (fail) {
3028 spin_lock_irq(&conf->device_lock);
3029 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
3030 conf->nr_queued++;
3031 spin_unlock_irq(&conf->device_lock);
3032 /*
3033 * In case freeze_array() is waiting for condition
3034 * nr_pending == nr_queued + extra to be true.
3035 */
3036 wake_up(&conf->wait_barrier);
3037 md_wakeup_thread(conf->mddev->thread);
3038 } else {
3039 if (test_bit(R10BIO_WriteError,
3040 &r10_bio->state))
3041 close_write(r10_bio);
3042 raid_end_bio_io(r10_bio);
3043 }
3044 }
3045 }
3046
raid10d(struct md_thread * thread)3047 static void raid10d(struct md_thread *thread)
3048 {
3049 struct mddev *mddev = thread->mddev;
3050 struct r10bio *r10_bio;
3051 unsigned long flags;
3052 struct r10conf *conf = mddev->private;
3053 struct list_head *head = &conf->retry_list;
3054 struct blk_plug plug;
3055
3056 md_check_recovery(mddev);
3057
3058 if (!list_empty_careful(&conf->bio_end_io_list) &&
3059 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3060 LIST_HEAD(tmp);
3061 spin_lock_irqsave(&conf->device_lock, flags);
3062 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3063 while (!list_empty(&conf->bio_end_io_list)) {
3064 list_move(conf->bio_end_io_list.prev, &tmp);
3065 conf->nr_queued--;
3066 }
3067 }
3068 spin_unlock_irqrestore(&conf->device_lock, flags);
3069 while (!list_empty(&tmp)) {
3070 r10_bio = list_first_entry(&tmp, struct r10bio,
3071 retry_list);
3072 list_del(&r10_bio->retry_list);
3073 if (mddev->degraded)
3074 set_bit(R10BIO_Degraded, &r10_bio->state);
3075
3076 if (test_bit(R10BIO_WriteError,
3077 &r10_bio->state))
3078 close_write(r10_bio);
3079 raid_end_bio_io(r10_bio);
3080 }
3081 }
3082
3083 blk_start_plug(&plug);
3084 for (;;) {
3085
3086 flush_pending_writes(conf);
3087
3088 spin_lock_irqsave(&conf->device_lock, flags);
3089 if (list_empty(head)) {
3090 spin_unlock_irqrestore(&conf->device_lock, flags);
3091 break;
3092 }
3093 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3094 list_del(head->prev);
3095 conf->nr_queued--;
3096 spin_unlock_irqrestore(&conf->device_lock, flags);
3097
3098 mddev = r10_bio->mddev;
3099 conf = mddev->private;
3100 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3101 test_bit(R10BIO_WriteError, &r10_bio->state))
3102 handle_write_completed(conf, r10_bio);
3103 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3104 reshape_request_write(mddev, r10_bio);
3105 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3106 sync_request_write(mddev, r10_bio);
3107 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3108 recovery_request_write(mddev, r10_bio);
3109 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3110 handle_read_error(mddev, r10_bio);
3111 else
3112 WARN_ON_ONCE(1);
3113
3114 cond_resched();
3115 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3116 md_check_recovery(mddev);
3117 }
3118 blk_finish_plug(&plug);
3119 }
3120
init_resync(struct r10conf * conf)3121 static int init_resync(struct r10conf *conf)
3122 {
3123 int ret, buffs, i;
3124
3125 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3126 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3127 conf->have_replacement = 0;
3128 for (i = 0; i < conf->geo.raid_disks; i++)
3129 if (conf->mirrors[i].replacement)
3130 conf->have_replacement = 1;
3131 ret = mempool_init(&conf->r10buf_pool, buffs,
3132 r10buf_pool_alloc, r10buf_pool_free, conf);
3133 if (ret)
3134 return ret;
3135 conf->next_resync = 0;
3136 return 0;
3137 }
3138
raid10_alloc_init_r10buf(struct r10conf * conf)3139 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3140 {
3141 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3142 struct rsync_pages *rp;
3143 struct bio *bio;
3144 int nalloc;
3145 int i;
3146
3147 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3148 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3149 nalloc = conf->copies; /* resync */
3150 else
3151 nalloc = 2; /* recovery */
3152
3153 for (i = 0; i < nalloc; i++) {
3154 bio = r10bio->devs[i].bio;
3155 rp = bio->bi_private;
3156 bio_reset(bio, NULL, 0);
3157 bio->bi_private = rp;
3158 bio = r10bio->devs[i].repl_bio;
3159 if (bio) {
3160 rp = bio->bi_private;
3161 bio_reset(bio, NULL, 0);
3162 bio->bi_private = rp;
3163 }
3164 }
3165 return r10bio;
3166 }
3167
3168 /*
3169 * Set cluster_sync_high since we need other nodes to add the
3170 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3171 */
raid10_set_cluster_sync_high(struct r10conf * conf)3172 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3173 {
3174 sector_t window_size;
3175 int extra_chunk, chunks;
3176
3177 /*
3178 * First, here we define "stripe" as a unit which across
3179 * all member devices one time, so we get chunks by use
3180 * raid_disks / near_copies. Otherwise, if near_copies is
3181 * close to raid_disks, then resync window could increases
3182 * linearly with the increase of raid_disks, which means
3183 * we will suspend a really large IO window while it is not
3184 * necessary. If raid_disks is not divisible by near_copies,
3185 * an extra chunk is needed to ensure the whole "stripe" is
3186 * covered.
3187 */
3188
3189 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3190 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3191 extra_chunk = 0;
3192 else
3193 extra_chunk = 1;
3194 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3195
3196 /*
3197 * At least use a 32M window to align with raid1's resync window
3198 */
3199 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3200 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3201
3202 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3203 }
3204
3205 /*
3206 * perform a "sync" on one "block"
3207 *
3208 * We need to make sure that no normal I/O request - particularly write
3209 * requests - conflict with active sync requests.
3210 *
3211 * This is achieved by tracking pending requests and a 'barrier' concept
3212 * that can be installed to exclude normal IO requests.
3213 *
3214 * Resync and recovery are handled very differently.
3215 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3216 *
3217 * For resync, we iterate over virtual addresses, read all copies,
3218 * and update if there are differences. If only one copy is live,
3219 * skip it.
3220 * For recovery, we iterate over physical addresses, read a good
3221 * value for each non-in_sync drive, and over-write.
3222 *
3223 * So, for recovery we may have several outstanding complex requests for a
3224 * given address, one for each out-of-sync device. We model this by allocating
3225 * a number of r10_bio structures, one for each out-of-sync device.
3226 * As we setup these structures, we collect all bio's together into a list
3227 * which we then process collectively to add pages, and then process again
3228 * to pass to submit_bio_noacct.
3229 *
3230 * The r10_bio structures are linked using a borrowed master_bio pointer.
3231 * This link is counted in ->remaining. When the r10_bio that points to NULL
3232 * has its remaining count decremented to 0, the whole complex operation
3233 * is complete.
3234 *
3235 */
3236
raid10_sync_request(struct mddev * mddev,sector_t sector_nr,int * skipped)3237 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3238 int *skipped)
3239 {
3240 struct r10conf *conf = mddev->private;
3241 struct r10bio *r10_bio;
3242 struct bio *biolist = NULL, *bio;
3243 sector_t max_sector, nr_sectors;
3244 int i;
3245 int max_sync;
3246 sector_t sync_blocks;
3247 sector_t sectors_skipped = 0;
3248 int chunks_skipped = 0;
3249 sector_t chunk_mask = conf->geo.chunk_mask;
3250 int page_idx = 0;
3251
3252 if (!mempool_initialized(&conf->r10buf_pool))
3253 if (init_resync(conf))
3254 return 0;
3255
3256 /*
3257 * Allow skipping a full rebuild for incremental assembly
3258 * of a clean array, like RAID1 does.
3259 */
3260 if (mddev->bitmap == NULL &&
3261 mddev->recovery_cp == MaxSector &&
3262 mddev->reshape_position == MaxSector &&
3263 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3264 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3265 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3266 conf->fullsync == 0) {
3267 *skipped = 1;
3268 return mddev->dev_sectors - sector_nr;
3269 }
3270
3271 skipped:
3272 max_sector = mddev->dev_sectors;
3273 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3274 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3275 max_sector = mddev->resync_max_sectors;
3276 if (sector_nr >= max_sector) {
3277 conf->cluster_sync_low = 0;
3278 conf->cluster_sync_high = 0;
3279
3280 /* If we aborted, we need to abort the
3281 * sync on the 'current' bitmap chucks (there can
3282 * be several when recovering multiple devices).
3283 * as we may have started syncing it but not finished.
3284 * We can find the current address in
3285 * mddev->curr_resync, but for recovery,
3286 * we need to convert that to several
3287 * virtual addresses.
3288 */
3289 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3290 end_reshape(conf);
3291 close_sync(conf);
3292 return 0;
3293 }
3294
3295 if (mddev->curr_resync < max_sector) { /* aborted */
3296 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3297 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3298 &sync_blocks, 1);
3299 else for (i = 0; i < conf->geo.raid_disks; i++) {
3300 sector_t sect =
3301 raid10_find_virt(conf, mddev->curr_resync, i);
3302 md_bitmap_end_sync(mddev->bitmap, sect,
3303 &sync_blocks, 1);
3304 }
3305 } else {
3306 /* completed sync */
3307 if ((!mddev->bitmap || conf->fullsync)
3308 && conf->have_replacement
3309 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3310 /* Completed a full sync so the replacements
3311 * are now fully recovered.
3312 */
3313 rcu_read_lock();
3314 for (i = 0; i < conf->geo.raid_disks; i++) {
3315 struct md_rdev *rdev =
3316 rcu_dereference(conf->mirrors[i].replacement);
3317 if (rdev)
3318 rdev->recovery_offset = MaxSector;
3319 }
3320 rcu_read_unlock();
3321 }
3322 conf->fullsync = 0;
3323 }
3324 md_bitmap_close_sync(mddev->bitmap);
3325 close_sync(conf);
3326 *skipped = 1;
3327 return sectors_skipped;
3328 }
3329
3330 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3331 return reshape_request(mddev, sector_nr, skipped);
3332
3333 if (chunks_skipped >= conf->geo.raid_disks) {
3334 /* if there has been nothing to do on any drive,
3335 * then there is nothing to do at all..
3336 */
3337 *skipped = 1;
3338 return (max_sector - sector_nr) + sectors_skipped;
3339 }
3340
3341 if (max_sector > mddev->resync_max)
3342 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3343
3344 /* make sure whole request will fit in a chunk - if chunks
3345 * are meaningful
3346 */
3347 if (conf->geo.near_copies < conf->geo.raid_disks &&
3348 max_sector > (sector_nr | chunk_mask))
3349 max_sector = (sector_nr | chunk_mask) + 1;
3350
3351 /*
3352 * If there is non-resync activity waiting for a turn, then let it
3353 * though before starting on this new sync request.
3354 */
3355 if (conf->nr_waiting)
3356 schedule_timeout_uninterruptible(1);
3357
3358 /* Again, very different code for resync and recovery.
3359 * Both must result in an r10bio with a list of bios that
3360 * have bi_end_io, bi_sector, bi_bdev set,
3361 * and bi_private set to the r10bio.
3362 * For recovery, we may actually create several r10bios
3363 * with 2 bios in each, that correspond to the bios in the main one.
3364 * In this case, the subordinate r10bios link back through a
3365 * borrowed master_bio pointer, and the counter in the master
3366 * includes a ref from each subordinate.
3367 */
3368 /* First, we decide what to do and set ->bi_end_io
3369 * To end_sync_read if we want to read, and
3370 * end_sync_write if we will want to write.
3371 */
3372
3373 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3374 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3375 /* recovery... the complicated one */
3376 int j;
3377 r10_bio = NULL;
3378
3379 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3380 int still_degraded;
3381 struct r10bio *rb2;
3382 sector_t sect;
3383 int must_sync;
3384 int any_working;
3385 int need_recover = 0;
3386 int need_replace = 0;
3387 struct raid10_info *mirror = &conf->mirrors[i];
3388 struct md_rdev *mrdev, *mreplace;
3389
3390 rcu_read_lock();
3391 mrdev = rcu_dereference(mirror->rdev);
3392 mreplace = rcu_dereference(mirror->replacement);
3393
3394 if (mrdev != NULL &&
3395 !test_bit(Faulty, &mrdev->flags) &&
3396 !test_bit(In_sync, &mrdev->flags))
3397 need_recover = 1;
3398 if (mreplace != NULL &&
3399 !test_bit(Faulty, &mreplace->flags))
3400 need_replace = 1;
3401
3402 if (!need_recover && !need_replace) {
3403 rcu_read_unlock();
3404 continue;
3405 }
3406
3407 still_degraded = 0;
3408 /* want to reconstruct this device */
3409 rb2 = r10_bio;
3410 sect = raid10_find_virt(conf, sector_nr, i);
3411 if (sect >= mddev->resync_max_sectors) {
3412 /* last stripe is not complete - don't
3413 * try to recover this sector.
3414 */
3415 rcu_read_unlock();
3416 continue;
3417 }
3418 if (mreplace && test_bit(Faulty, &mreplace->flags))
3419 mreplace = NULL;
3420 /* Unless we are doing a full sync, or a replacement
3421 * we only need to recover the block if it is set in
3422 * the bitmap
3423 */
3424 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3425 &sync_blocks, 1);
3426 if (sync_blocks < max_sync)
3427 max_sync = sync_blocks;
3428 if (!must_sync &&
3429 mreplace == NULL &&
3430 !conf->fullsync) {
3431 /* yep, skip the sync_blocks here, but don't assume
3432 * that there will never be anything to do here
3433 */
3434 chunks_skipped = -1;
3435 rcu_read_unlock();
3436 continue;
3437 }
3438 atomic_inc(&mrdev->nr_pending);
3439 if (mreplace)
3440 atomic_inc(&mreplace->nr_pending);
3441 rcu_read_unlock();
3442
3443 r10_bio = raid10_alloc_init_r10buf(conf);
3444 r10_bio->state = 0;
3445 raise_barrier(conf, rb2 != NULL);
3446 atomic_set(&r10_bio->remaining, 0);
3447
3448 r10_bio->master_bio = (struct bio*)rb2;
3449 if (rb2)
3450 atomic_inc(&rb2->remaining);
3451 r10_bio->mddev = mddev;
3452 set_bit(R10BIO_IsRecover, &r10_bio->state);
3453 r10_bio->sector = sect;
3454
3455 raid10_find_phys(conf, r10_bio);
3456
3457 /* Need to check if the array will still be
3458 * degraded
3459 */
3460 rcu_read_lock();
3461 for (j = 0; j < conf->geo.raid_disks; j++) {
3462 struct md_rdev *rdev = rcu_dereference(
3463 conf->mirrors[j].rdev);
3464 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3465 still_degraded = 1;
3466 break;
3467 }
3468 }
3469
3470 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3471 &sync_blocks, still_degraded);
3472
3473 any_working = 0;
3474 for (j=0; j<conf->copies;j++) {
3475 int k;
3476 int d = r10_bio->devs[j].devnum;
3477 sector_t from_addr, to_addr;
3478 struct md_rdev *rdev =
3479 rcu_dereference(conf->mirrors[d].rdev);
3480 sector_t sector, first_bad;
3481 int bad_sectors;
3482 if (!rdev ||
3483 !test_bit(In_sync, &rdev->flags))
3484 continue;
3485 /* This is where we read from */
3486 any_working = 1;
3487 sector = r10_bio->devs[j].addr;
3488
3489 if (is_badblock(rdev, sector, max_sync,
3490 &first_bad, &bad_sectors)) {
3491 if (first_bad > sector)
3492 max_sync = first_bad - sector;
3493 else {
3494 bad_sectors -= (sector
3495 - first_bad);
3496 if (max_sync > bad_sectors)
3497 max_sync = bad_sectors;
3498 continue;
3499 }
3500 }
3501 bio = r10_bio->devs[0].bio;
3502 bio->bi_next = biolist;
3503 biolist = bio;
3504 bio->bi_end_io = end_sync_read;
3505 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3506 if (test_bit(FailFast, &rdev->flags))
3507 bio->bi_opf |= MD_FAILFAST;
3508 from_addr = r10_bio->devs[j].addr;
3509 bio->bi_iter.bi_sector = from_addr +
3510 rdev->data_offset;
3511 bio_set_dev(bio, rdev->bdev);
3512 atomic_inc(&rdev->nr_pending);
3513 /* and we write to 'i' (if not in_sync) */
3514
3515 for (k=0; k<conf->copies; k++)
3516 if (r10_bio->devs[k].devnum == i)
3517 break;
3518 BUG_ON(k == conf->copies);
3519 to_addr = r10_bio->devs[k].addr;
3520 r10_bio->devs[0].devnum = d;
3521 r10_bio->devs[0].addr = from_addr;
3522 r10_bio->devs[1].devnum = i;
3523 r10_bio->devs[1].addr = to_addr;
3524
3525 if (need_recover) {
3526 bio = r10_bio->devs[1].bio;
3527 bio->bi_next = biolist;
3528 biolist = bio;
3529 bio->bi_end_io = end_sync_write;
3530 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3531 bio->bi_iter.bi_sector = to_addr
3532 + mrdev->data_offset;
3533 bio_set_dev(bio, mrdev->bdev);
3534 atomic_inc(&r10_bio->remaining);
3535 } else
3536 r10_bio->devs[1].bio->bi_end_io = NULL;
3537
3538 /* and maybe write to replacement */
3539 bio = r10_bio->devs[1].repl_bio;
3540 if (bio)
3541 bio->bi_end_io = NULL;
3542 /* Note: if need_replace, then bio
3543 * cannot be NULL as r10buf_pool_alloc will
3544 * have allocated it.
3545 */
3546 if (!need_replace)
3547 break;
3548 bio->bi_next = biolist;
3549 biolist = bio;
3550 bio->bi_end_io = end_sync_write;
3551 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3552 bio->bi_iter.bi_sector = to_addr +
3553 mreplace->data_offset;
3554 bio_set_dev(bio, mreplace->bdev);
3555 atomic_inc(&r10_bio->remaining);
3556 break;
3557 }
3558 rcu_read_unlock();
3559 if (j == conf->copies) {
3560 /* Cannot recover, so abort the recovery or
3561 * record a bad block */
3562 if (any_working) {
3563 /* problem is that there are bad blocks
3564 * on other device(s)
3565 */
3566 int k;
3567 for (k = 0; k < conf->copies; k++)
3568 if (r10_bio->devs[k].devnum == i)
3569 break;
3570 if (!test_bit(In_sync,
3571 &mrdev->flags)
3572 && !rdev_set_badblocks(
3573 mrdev,
3574 r10_bio->devs[k].addr,
3575 max_sync, 0))
3576 any_working = 0;
3577 if (mreplace &&
3578 !rdev_set_badblocks(
3579 mreplace,
3580 r10_bio->devs[k].addr,
3581 max_sync, 0))
3582 any_working = 0;
3583 }
3584 if (!any_working) {
3585 if (!test_and_set_bit(MD_RECOVERY_INTR,
3586 &mddev->recovery))
3587 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3588 mdname(mddev));
3589 mirror->recovery_disabled
3590 = mddev->recovery_disabled;
3591 }
3592 put_buf(r10_bio);
3593 if (rb2)
3594 atomic_dec(&rb2->remaining);
3595 r10_bio = rb2;
3596 rdev_dec_pending(mrdev, mddev);
3597 if (mreplace)
3598 rdev_dec_pending(mreplace, mddev);
3599 break;
3600 }
3601 rdev_dec_pending(mrdev, mddev);
3602 if (mreplace)
3603 rdev_dec_pending(mreplace, mddev);
3604 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3605 /* Only want this if there is elsewhere to
3606 * read from. 'j' is currently the first
3607 * readable copy.
3608 */
3609 int targets = 1;
3610 for (; j < conf->copies; j++) {
3611 int d = r10_bio->devs[j].devnum;
3612 if (conf->mirrors[d].rdev &&
3613 test_bit(In_sync,
3614 &conf->mirrors[d].rdev->flags))
3615 targets++;
3616 }
3617 if (targets == 1)
3618 r10_bio->devs[0].bio->bi_opf
3619 &= ~MD_FAILFAST;
3620 }
3621 }
3622 if (biolist == NULL) {
3623 while (r10_bio) {
3624 struct r10bio *rb2 = r10_bio;
3625 r10_bio = (struct r10bio*) rb2->master_bio;
3626 rb2->master_bio = NULL;
3627 put_buf(rb2);
3628 }
3629 goto giveup;
3630 }
3631 } else {
3632 /* resync. Schedule a read for every block at this virt offset */
3633 int count = 0;
3634
3635 /*
3636 * Since curr_resync_completed could probably not update in
3637 * time, and we will set cluster_sync_low based on it.
3638 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3639 * safety reason, which ensures curr_resync_completed is
3640 * updated in bitmap_cond_end_sync.
3641 */
3642 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3643 mddev_is_clustered(mddev) &&
3644 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3645
3646 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3647 &sync_blocks, mddev->degraded) &&
3648 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3649 &mddev->recovery)) {
3650 /* We can skip this block */
3651 *skipped = 1;
3652 return sync_blocks + sectors_skipped;
3653 }
3654 if (sync_blocks < max_sync)
3655 max_sync = sync_blocks;
3656 r10_bio = raid10_alloc_init_r10buf(conf);
3657 r10_bio->state = 0;
3658
3659 r10_bio->mddev = mddev;
3660 atomic_set(&r10_bio->remaining, 0);
3661 raise_barrier(conf, 0);
3662 conf->next_resync = sector_nr;
3663
3664 r10_bio->master_bio = NULL;
3665 r10_bio->sector = sector_nr;
3666 set_bit(R10BIO_IsSync, &r10_bio->state);
3667 raid10_find_phys(conf, r10_bio);
3668 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3669
3670 for (i = 0; i < conf->copies; i++) {
3671 int d = r10_bio->devs[i].devnum;
3672 sector_t first_bad, sector;
3673 int bad_sectors;
3674 struct md_rdev *rdev;
3675
3676 if (r10_bio->devs[i].repl_bio)
3677 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3678
3679 bio = r10_bio->devs[i].bio;
3680 bio->bi_status = BLK_STS_IOERR;
3681 rcu_read_lock();
3682 rdev = rcu_dereference(conf->mirrors[d].rdev);
3683 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3684 rcu_read_unlock();
3685 continue;
3686 }
3687 sector = r10_bio->devs[i].addr;
3688 if (is_badblock(rdev, sector, max_sync,
3689 &first_bad, &bad_sectors)) {
3690 if (first_bad > sector)
3691 max_sync = first_bad - sector;
3692 else {
3693 bad_sectors -= (sector - first_bad);
3694 if (max_sync > bad_sectors)
3695 max_sync = bad_sectors;
3696 rcu_read_unlock();
3697 continue;
3698 }
3699 }
3700 atomic_inc(&rdev->nr_pending);
3701 atomic_inc(&r10_bio->remaining);
3702 bio->bi_next = biolist;
3703 biolist = bio;
3704 bio->bi_end_io = end_sync_read;
3705 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3706 if (test_bit(FailFast, &rdev->flags))
3707 bio->bi_opf |= MD_FAILFAST;
3708 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3709 bio_set_dev(bio, rdev->bdev);
3710 count++;
3711
3712 rdev = rcu_dereference(conf->mirrors[d].replacement);
3713 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3714 rcu_read_unlock();
3715 continue;
3716 }
3717 atomic_inc(&rdev->nr_pending);
3718
3719 /* Need to set up for writing to the replacement */
3720 bio = r10_bio->devs[i].repl_bio;
3721 bio->bi_status = BLK_STS_IOERR;
3722
3723 sector = r10_bio->devs[i].addr;
3724 bio->bi_next = biolist;
3725 biolist = bio;
3726 bio->bi_end_io = end_sync_write;
3727 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3728 if (test_bit(FailFast, &rdev->flags))
3729 bio->bi_opf |= MD_FAILFAST;
3730 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3731 bio_set_dev(bio, rdev->bdev);
3732 count++;
3733 rcu_read_unlock();
3734 }
3735
3736 if (count < 2) {
3737 for (i=0; i<conf->copies; i++) {
3738 int d = r10_bio->devs[i].devnum;
3739 if (r10_bio->devs[i].bio->bi_end_io)
3740 rdev_dec_pending(conf->mirrors[d].rdev,
3741 mddev);
3742 if (r10_bio->devs[i].repl_bio &&
3743 r10_bio->devs[i].repl_bio->bi_end_io)
3744 rdev_dec_pending(
3745 conf->mirrors[d].replacement,
3746 mddev);
3747 }
3748 put_buf(r10_bio);
3749 biolist = NULL;
3750 goto giveup;
3751 }
3752 }
3753
3754 nr_sectors = 0;
3755 if (sector_nr + max_sync < max_sector)
3756 max_sector = sector_nr + max_sync;
3757 do {
3758 struct page *page;
3759 int len = PAGE_SIZE;
3760 if (sector_nr + (len>>9) > max_sector)
3761 len = (max_sector - sector_nr) << 9;
3762 if (len == 0)
3763 break;
3764 for (bio= biolist ; bio ; bio=bio->bi_next) {
3765 struct resync_pages *rp = get_resync_pages(bio);
3766 page = resync_fetch_page(rp, page_idx);
3767 /*
3768 * won't fail because the vec table is big enough
3769 * to hold all these pages
3770 */
3771 bio_add_page(bio, page, len, 0);
3772 }
3773 nr_sectors += len>>9;
3774 sector_nr += len>>9;
3775 } while (++page_idx < RESYNC_PAGES);
3776 r10_bio->sectors = nr_sectors;
3777
3778 if (mddev_is_clustered(mddev) &&
3779 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3780 /* It is resync not recovery */
3781 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3782 conf->cluster_sync_low = mddev->curr_resync_completed;
3783 raid10_set_cluster_sync_high(conf);
3784 /* Send resync message */
3785 md_cluster_ops->resync_info_update(mddev,
3786 conf->cluster_sync_low,
3787 conf->cluster_sync_high);
3788 }
3789 } else if (mddev_is_clustered(mddev)) {
3790 /* This is recovery not resync */
3791 sector_t sect_va1, sect_va2;
3792 bool broadcast_msg = false;
3793
3794 for (i = 0; i < conf->geo.raid_disks; i++) {
3795 /*
3796 * sector_nr is a device address for recovery, so we
3797 * need translate it to array address before compare
3798 * with cluster_sync_high.
3799 */
3800 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3801
3802 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3803 broadcast_msg = true;
3804 /*
3805 * curr_resync_completed is similar as
3806 * sector_nr, so make the translation too.
3807 */
3808 sect_va2 = raid10_find_virt(conf,
3809 mddev->curr_resync_completed, i);
3810
3811 if (conf->cluster_sync_low == 0 ||
3812 conf->cluster_sync_low > sect_va2)
3813 conf->cluster_sync_low = sect_va2;
3814 }
3815 }
3816 if (broadcast_msg) {
3817 raid10_set_cluster_sync_high(conf);
3818 md_cluster_ops->resync_info_update(mddev,
3819 conf->cluster_sync_low,
3820 conf->cluster_sync_high);
3821 }
3822 }
3823
3824 while (biolist) {
3825 bio = biolist;
3826 biolist = biolist->bi_next;
3827
3828 bio->bi_next = NULL;
3829 r10_bio = get_resync_r10bio(bio);
3830 r10_bio->sectors = nr_sectors;
3831
3832 if (bio->bi_end_io == end_sync_read) {
3833 md_sync_acct_bio(bio, nr_sectors);
3834 bio->bi_status = 0;
3835 submit_bio_noacct(bio);
3836 }
3837 }
3838
3839 if (sectors_skipped)
3840 /* pretend they weren't skipped, it makes
3841 * no important difference in this case
3842 */
3843 md_done_sync(mddev, sectors_skipped, 1);
3844
3845 return sectors_skipped + nr_sectors;
3846 giveup:
3847 /* There is nowhere to write, so all non-sync
3848 * drives must be failed or in resync, all drives
3849 * have a bad block, so try the next chunk...
3850 */
3851 if (sector_nr + max_sync < max_sector)
3852 max_sector = sector_nr + max_sync;
3853
3854 sectors_skipped += (max_sector - sector_nr);
3855 chunks_skipped ++;
3856 sector_nr = max_sector;
3857 goto skipped;
3858 }
3859
3860 static sector_t
raid10_size(struct mddev * mddev,sector_t sectors,int raid_disks)3861 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3862 {
3863 sector_t size;
3864 struct r10conf *conf = mddev->private;
3865
3866 if (!raid_disks)
3867 raid_disks = min(conf->geo.raid_disks,
3868 conf->prev.raid_disks);
3869 if (!sectors)
3870 sectors = conf->dev_sectors;
3871
3872 size = sectors >> conf->geo.chunk_shift;
3873 sector_div(size, conf->geo.far_copies);
3874 size = size * raid_disks;
3875 sector_div(size, conf->geo.near_copies);
3876
3877 return size << conf->geo.chunk_shift;
3878 }
3879
calc_sectors(struct r10conf * conf,sector_t size)3880 static void calc_sectors(struct r10conf *conf, sector_t size)
3881 {
3882 /* Calculate the number of sectors-per-device that will
3883 * actually be used, and set conf->dev_sectors and
3884 * conf->stride
3885 */
3886
3887 size = size >> conf->geo.chunk_shift;
3888 sector_div(size, conf->geo.far_copies);
3889 size = size * conf->geo.raid_disks;
3890 sector_div(size, conf->geo.near_copies);
3891 /* 'size' is now the number of chunks in the array */
3892 /* calculate "used chunks per device" */
3893 size = size * conf->copies;
3894
3895 /* We need to round up when dividing by raid_disks to
3896 * get the stride size.
3897 */
3898 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3899
3900 conf->dev_sectors = size << conf->geo.chunk_shift;
3901
3902 if (conf->geo.far_offset)
3903 conf->geo.stride = 1 << conf->geo.chunk_shift;
3904 else {
3905 sector_div(size, conf->geo.far_copies);
3906 conf->geo.stride = size << conf->geo.chunk_shift;
3907 }
3908 }
3909
3910 enum geo_type {geo_new, geo_old, geo_start};
setup_geo(struct geom * geo,struct mddev * mddev,enum geo_type new)3911 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3912 {
3913 int nc, fc, fo;
3914 int layout, chunk, disks;
3915 switch (new) {
3916 case geo_old:
3917 layout = mddev->layout;
3918 chunk = mddev->chunk_sectors;
3919 disks = mddev->raid_disks - mddev->delta_disks;
3920 break;
3921 case geo_new:
3922 layout = mddev->new_layout;
3923 chunk = mddev->new_chunk_sectors;
3924 disks = mddev->raid_disks;
3925 break;
3926 default: /* avoid 'may be unused' warnings */
3927 case geo_start: /* new when starting reshape - raid_disks not
3928 * updated yet. */
3929 layout = mddev->new_layout;
3930 chunk = mddev->new_chunk_sectors;
3931 disks = mddev->raid_disks + mddev->delta_disks;
3932 break;
3933 }
3934 if (layout >> 19)
3935 return -1;
3936 if (chunk < (PAGE_SIZE >> 9) ||
3937 !is_power_of_2(chunk))
3938 return -2;
3939 nc = layout & 255;
3940 fc = (layout >> 8) & 255;
3941 fo = layout & (1<<16);
3942 geo->raid_disks = disks;
3943 geo->near_copies = nc;
3944 geo->far_copies = fc;
3945 geo->far_offset = fo;
3946 switch (layout >> 17) {
3947 case 0: /* original layout. simple but not always optimal */
3948 geo->far_set_size = disks;
3949 break;
3950 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3951 * actually using this, but leave code here just in case.*/
3952 geo->far_set_size = disks/fc;
3953 WARN(geo->far_set_size < fc,
3954 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3955 break;
3956 case 2: /* "improved" layout fixed to match documentation */
3957 geo->far_set_size = fc * nc;
3958 break;
3959 default: /* Not a valid layout */
3960 return -1;
3961 }
3962 geo->chunk_mask = chunk - 1;
3963 geo->chunk_shift = ffz(~chunk);
3964 return nc*fc;
3965 }
3966
setup_conf(struct mddev * mddev)3967 static struct r10conf *setup_conf(struct mddev *mddev)
3968 {
3969 struct r10conf *conf = NULL;
3970 int err = -EINVAL;
3971 struct geom geo;
3972 int copies;
3973
3974 copies = setup_geo(&geo, mddev, geo_new);
3975
3976 if (copies == -2) {
3977 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3978 mdname(mddev), PAGE_SIZE);
3979 goto out;
3980 }
3981
3982 if (copies < 2 || copies > mddev->raid_disks) {
3983 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3984 mdname(mddev), mddev->new_layout);
3985 goto out;
3986 }
3987
3988 err = -ENOMEM;
3989 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3990 if (!conf)
3991 goto out;
3992
3993 /* FIXME calc properly */
3994 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3995 sizeof(struct raid10_info),
3996 GFP_KERNEL);
3997 if (!conf->mirrors)
3998 goto out;
3999
4000 conf->tmppage = alloc_page(GFP_KERNEL);
4001 if (!conf->tmppage)
4002 goto out;
4003
4004 conf->geo = geo;
4005 conf->copies = copies;
4006 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
4007 rbio_pool_free, conf);
4008 if (err)
4009 goto out;
4010
4011 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
4012 if (err)
4013 goto out;
4014
4015 calc_sectors(conf, mddev->dev_sectors);
4016 if (mddev->reshape_position == MaxSector) {
4017 conf->prev = conf->geo;
4018 conf->reshape_progress = MaxSector;
4019 } else {
4020 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
4021 err = -EINVAL;
4022 goto out;
4023 }
4024 conf->reshape_progress = mddev->reshape_position;
4025 if (conf->prev.far_offset)
4026 conf->prev.stride = 1 << conf->prev.chunk_shift;
4027 else
4028 /* far_copies must be 1 */
4029 conf->prev.stride = conf->dev_sectors;
4030 }
4031 conf->reshape_safe = conf->reshape_progress;
4032 spin_lock_init(&conf->device_lock);
4033 INIT_LIST_HEAD(&conf->retry_list);
4034 INIT_LIST_HEAD(&conf->bio_end_io_list);
4035
4036 spin_lock_init(&conf->resync_lock);
4037 init_waitqueue_head(&conf->wait_barrier);
4038 atomic_set(&conf->nr_pending, 0);
4039
4040 err = -ENOMEM;
4041 conf->thread = md_register_thread(raid10d, mddev, "raid10");
4042 if (!conf->thread)
4043 goto out;
4044
4045 conf->mddev = mddev;
4046 return conf;
4047
4048 out:
4049 if (conf) {
4050 mempool_exit(&conf->r10bio_pool);
4051 kfree(conf->mirrors);
4052 safe_put_page(conf->tmppage);
4053 bioset_exit(&conf->bio_split);
4054 kfree(conf);
4055 }
4056 return ERR_PTR(err);
4057 }
4058
raid10_set_io_opt(struct r10conf * conf)4059 static void raid10_set_io_opt(struct r10conf *conf)
4060 {
4061 int raid_disks = conf->geo.raid_disks;
4062
4063 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4064 raid_disks /= conf->geo.near_copies;
4065 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4066 raid_disks);
4067 }
4068
raid10_run(struct mddev * mddev)4069 static int raid10_run(struct mddev *mddev)
4070 {
4071 struct r10conf *conf;
4072 int i, disk_idx;
4073 struct raid10_info *disk;
4074 struct md_rdev *rdev;
4075 sector_t size;
4076 sector_t min_offset_diff = 0;
4077 int first = 1;
4078
4079 if (mddev_init_writes_pending(mddev) < 0)
4080 return -ENOMEM;
4081
4082 if (mddev->private == NULL) {
4083 conf = setup_conf(mddev);
4084 if (IS_ERR(conf))
4085 return PTR_ERR(conf);
4086 mddev->private = conf;
4087 }
4088 conf = mddev->private;
4089 if (!conf)
4090 goto out;
4091
4092 if (mddev_is_clustered(conf->mddev)) {
4093 int fc, fo;
4094
4095 fc = (mddev->layout >> 8) & 255;
4096 fo = mddev->layout & (1<<16);
4097 if (fc > 1 || fo > 0) {
4098 pr_err("only near layout is supported by clustered"
4099 " raid10\n");
4100 goto out_free_conf;
4101 }
4102 }
4103
4104 mddev->thread = conf->thread;
4105 conf->thread = NULL;
4106
4107 if (mddev->queue) {
4108 blk_queue_max_discard_sectors(mddev->queue,
4109 UINT_MAX);
4110 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4111 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4112 raid10_set_io_opt(conf);
4113 }
4114
4115 rdev_for_each(rdev, mddev) {
4116 long long diff;
4117
4118 disk_idx = rdev->raid_disk;
4119 if (disk_idx < 0)
4120 continue;
4121 if (disk_idx >= conf->geo.raid_disks &&
4122 disk_idx >= conf->prev.raid_disks)
4123 continue;
4124 disk = conf->mirrors + disk_idx;
4125
4126 if (test_bit(Replacement, &rdev->flags)) {
4127 if (disk->replacement)
4128 goto out_free_conf;
4129 disk->replacement = rdev;
4130 } else {
4131 if (disk->rdev)
4132 goto out_free_conf;
4133 disk->rdev = rdev;
4134 }
4135 diff = (rdev->new_data_offset - rdev->data_offset);
4136 if (!mddev->reshape_backwards)
4137 diff = -diff;
4138 if (diff < 0)
4139 diff = 0;
4140 if (first || diff < min_offset_diff)
4141 min_offset_diff = diff;
4142
4143 if (mddev->gendisk)
4144 disk_stack_limits(mddev->gendisk, rdev->bdev,
4145 rdev->data_offset << 9);
4146
4147 disk->head_position = 0;
4148 first = 0;
4149 }
4150
4151 /* need to check that every block has at least one working mirror */
4152 if (!enough(conf, -1)) {
4153 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4154 mdname(mddev));
4155 goto out_free_conf;
4156 }
4157
4158 if (conf->reshape_progress != MaxSector) {
4159 /* must ensure that shape change is supported */
4160 if (conf->geo.far_copies != 1 &&
4161 conf->geo.far_offset == 0)
4162 goto out_free_conf;
4163 if (conf->prev.far_copies != 1 &&
4164 conf->prev.far_offset == 0)
4165 goto out_free_conf;
4166 }
4167
4168 mddev->degraded = 0;
4169 for (i = 0;
4170 i < conf->geo.raid_disks
4171 || i < conf->prev.raid_disks;
4172 i++) {
4173
4174 disk = conf->mirrors + i;
4175
4176 if (!disk->rdev && disk->replacement) {
4177 /* The replacement is all we have - use it */
4178 disk->rdev = disk->replacement;
4179 disk->replacement = NULL;
4180 clear_bit(Replacement, &disk->rdev->flags);
4181 }
4182
4183 if (!disk->rdev ||
4184 !test_bit(In_sync, &disk->rdev->flags)) {
4185 disk->head_position = 0;
4186 mddev->degraded++;
4187 if (disk->rdev &&
4188 disk->rdev->saved_raid_disk < 0)
4189 conf->fullsync = 1;
4190 }
4191
4192 if (disk->replacement &&
4193 !test_bit(In_sync, &disk->replacement->flags) &&
4194 disk->replacement->saved_raid_disk < 0) {
4195 conf->fullsync = 1;
4196 }
4197
4198 disk->recovery_disabled = mddev->recovery_disabled - 1;
4199 }
4200
4201 if (mddev->recovery_cp != MaxSector)
4202 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4203 mdname(mddev));
4204 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4205 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4206 conf->geo.raid_disks);
4207 /*
4208 * Ok, everything is just fine now
4209 */
4210 mddev->dev_sectors = conf->dev_sectors;
4211 size = raid10_size(mddev, 0, 0);
4212 md_set_array_sectors(mddev, size);
4213 mddev->resync_max_sectors = size;
4214 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4215
4216 if (md_integrity_register(mddev))
4217 goto out_free_conf;
4218
4219 if (conf->reshape_progress != MaxSector) {
4220 unsigned long before_length, after_length;
4221
4222 before_length = ((1 << conf->prev.chunk_shift) *
4223 conf->prev.far_copies);
4224 after_length = ((1 << conf->geo.chunk_shift) *
4225 conf->geo.far_copies);
4226
4227 if (max(before_length, after_length) > min_offset_diff) {
4228 /* This cannot work */
4229 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4230 goto out_free_conf;
4231 }
4232 conf->offset_diff = min_offset_diff;
4233
4234 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4235 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4236 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4237 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4238 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4239 "reshape");
4240 if (!mddev->sync_thread)
4241 goto out_free_conf;
4242 }
4243
4244 return 0;
4245
4246 out_free_conf:
4247 md_unregister_thread(&mddev->thread);
4248 mempool_exit(&conf->r10bio_pool);
4249 safe_put_page(conf->tmppage);
4250 kfree(conf->mirrors);
4251 kfree(conf);
4252 mddev->private = NULL;
4253 out:
4254 return -EIO;
4255 }
4256
raid10_free(struct mddev * mddev,void * priv)4257 static void raid10_free(struct mddev *mddev, void *priv)
4258 {
4259 struct r10conf *conf = priv;
4260
4261 mempool_exit(&conf->r10bio_pool);
4262 safe_put_page(conf->tmppage);
4263 kfree(conf->mirrors);
4264 kfree(conf->mirrors_old);
4265 kfree(conf->mirrors_new);
4266 bioset_exit(&conf->bio_split);
4267 kfree(conf);
4268 }
4269
raid10_quiesce(struct mddev * mddev,int quiesce)4270 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4271 {
4272 struct r10conf *conf = mddev->private;
4273
4274 if (quiesce)
4275 raise_barrier(conf, 0);
4276 else
4277 lower_barrier(conf);
4278 }
4279
raid10_resize(struct mddev * mddev,sector_t sectors)4280 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4281 {
4282 /* Resize of 'far' arrays is not supported.
4283 * For 'near' and 'offset' arrays we can set the
4284 * number of sectors used to be an appropriate multiple
4285 * of the chunk size.
4286 * For 'offset', this is far_copies*chunksize.
4287 * For 'near' the multiplier is the LCM of
4288 * near_copies and raid_disks.
4289 * So if far_copies > 1 && !far_offset, fail.
4290 * Else find LCM(raid_disks, near_copy)*far_copies and
4291 * multiply by chunk_size. Then round to this number.
4292 * This is mostly done by raid10_size()
4293 */
4294 struct r10conf *conf = mddev->private;
4295 sector_t oldsize, size;
4296
4297 if (mddev->reshape_position != MaxSector)
4298 return -EBUSY;
4299
4300 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4301 return -EINVAL;
4302
4303 oldsize = raid10_size(mddev, 0, 0);
4304 size = raid10_size(mddev, sectors, 0);
4305 if (mddev->external_size &&
4306 mddev->array_sectors > size)
4307 return -EINVAL;
4308 if (mddev->bitmap) {
4309 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4310 if (ret)
4311 return ret;
4312 }
4313 md_set_array_sectors(mddev, size);
4314 if (sectors > mddev->dev_sectors &&
4315 mddev->recovery_cp > oldsize) {
4316 mddev->recovery_cp = oldsize;
4317 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4318 }
4319 calc_sectors(conf, sectors);
4320 mddev->dev_sectors = conf->dev_sectors;
4321 mddev->resync_max_sectors = size;
4322 return 0;
4323 }
4324
raid10_takeover_raid0(struct mddev * mddev,sector_t size,int devs)4325 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4326 {
4327 struct md_rdev *rdev;
4328 struct r10conf *conf;
4329
4330 if (mddev->degraded > 0) {
4331 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4332 mdname(mddev));
4333 return ERR_PTR(-EINVAL);
4334 }
4335 sector_div(size, devs);
4336
4337 /* Set new parameters */
4338 mddev->new_level = 10;
4339 /* new layout: far_copies = 1, near_copies = 2 */
4340 mddev->new_layout = (1<<8) + 2;
4341 mddev->new_chunk_sectors = mddev->chunk_sectors;
4342 mddev->delta_disks = mddev->raid_disks;
4343 mddev->raid_disks *= 2;
4344 /* make sure it will be not marked as dirty */
4345 mddev->recovery_cp = MaxSector;
4346 mddev->dev_sectors = size;
4347
4348 conf = setup_conf(mddev);
4349 if (!IS_ERR(conf)) {
4350 rdev_for_each(rdev, mddev)
4351 if (rdev->raid_disk >= 0) {
4352 rdev->new_raid_disk = rdev->raid_disk * 2;
4353 rdev->sectors = size;
4354 }
4355 conf->barrier = 1;
4356 }
4357
4358 return conf;
4359 }
4360
raid10_takeover(struct mddev * mddev)4361 static void *raid10_takeover(struct mddev *mddev)
4362 {
4363 struct r0conf *raid0_conf;
4364
4365 /* raid10 can take over:
4366 * raid0 - providing it has only two drives
4367 */
4368 if (mddev->level == 0) {
4369 /* for raid0 takeover only one zone is supported */
4370 raid0_conf = mddev->private;
4371 if (raid0_conf->nr_strip_zones > 1) {
4372 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4373 mdname(mddev));
4374 return ERR_PTR(-EINVAL);
4375 }
4376 return raid10_takeover_raid0(mddev,
4377 raid0_conf->strip_zone->zone_end,
4378 raid0_conf->strip_zone->nb_dev);
4379 }
4380 return ERR_PTR(-EINVAL);
4381 }
4382
raid10_check_reshape(struct mddev * mddev)4383 static int raid10_check_reshape(struct mddev *mddev)
4384 {
4385 /* Called when there is a request to change
4386 * - layout (to ->new_layout)
4387 * - chunk size (to ->new_chunk_sectors)
4388 * - raid_disks (by delta_disks)
4389 * or when trying to restart a reshape that was ongoing.
4390 *
4391 * We need to validate the request and possibly allocate
4392 * space if that might be an issue later.
4393 *
4394 * Currently we reject any reshape of a 'far' mode array,
4395 * allow chunk size to change if new is generally acceptable,
4396 * allow raid_disks to increase, and allow
4397 * a switch between 'near' mode and 'offset' mode.
4398 */
4399 struct r10conf *conf = mddev->private;
4400 struct geom geo;
4401
4402 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4403 return -EINVAL;
4404
4405 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4406 /* mustn't change number of copies */
4407 return -EINVAL;
4408 if (geo.far_copies > 1 && !geo.far_offset)
4409 /* Cannot switch to 'far' mode */
4410 return -EINVAL;
4411
4412 if (mddev->array_sectors & geo.chunk_mask)
4413 /* not factor of array size */
4414 return -EINVAL;
4415
4416 if (!enough(conf, -1))
4417 return -EINVAL;
4418
4419 kfree(conf->mirrors_new);
4420 conf->mirrors_new = NULL;
4421 if (mddev->delta_disks > 0) {
4422 /* allocate new 'mirrors' list */
4423 conf->mirrors_new =
4424 kcalloc(mddev->raid_disks + mddev->delta_disks,
4425 sizeof(struct raid10_info),
4426 GFP_KERNEL);
4427 if (!conf->mirrors_new)
4428 return -ENOMEM;
4429 }
4430 return 0;
4431 }
4432
4433 /*
4434 * Need to check if array has failed when deciding whether to:
4435 * - start an array
4436 * - remove non-faulty devices
4437 * - add a spare
4438 * - allow a reshape
4439 * This determination is simple when no reshape is happening.
4440 * However if there is a reshape, we need to carefully check
4441 * both the before and after sections.
4442 * This is because some failed devices may only affect one
4443 * of the two sections, and some non-in_sync devices may
4444 * be insync in the section most affected by failed devices.
4445 */
calc_degraded(struct r10conf * conf)4446 static int calc_degraded(struct r10conf *conf)
4447 {
4448 int degraded, degraded2;
4449 int i;
4450
4451 rcu_read_lock();
4452 degraded = 0;
4453 /* 'prev' section first */
4454 for (i = 0; i < conf->prev.raid_disks; i++) {
4455 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4456 if (!rdev || test_bit(Faulty, &rdev->flags))
4457 degraded++;
4458 else if (!test_bit(In_sync, &rdev->flags))
4459 /* When we can reduce the number of devices in
4460 * an array, this might not contribute to
4461 * 'degraded'. It does now.
4462 */
4463 degraded++;
4464 }
4465 rcu_read_unlock();
4466 if (conf->geo.raid_disks == conf->prev.raid_disks)
4467 return degraded;
4468 rcu_read_lock();
4469 degraded2 = 0;
4470 for (i = 0; i < conf->geo.raid_disks; i++) {
4471 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4472 if (!rdev || test_bit(Faulty, &rdev->flags))
4473 degraded2++;
4474 else if (!test_bit(In_sync, &rdev->flags)) {
4475 /* If reshape is increasing the number of devices,
4476 * this section has already been recovered, so
4477 * it doesn't contribute to degraded.
4478 * else it does.
4479 */
4480 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4481 degraded2++;
4482 }
4483 }
4484 rcu_read_unlock();
4485 if (degraded2 > degraded)
4486 return degraded2;
4487 return degraded;
4488 }
4489
raid10_start_reshape(struct mddev * mddev)4490 static int raid10_start_reshape(struct mddev *mddev)
4491 {
4492 /* A 'reshape' has been requested. This commits
4493 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4494 * This also checks if there are enough spares and adds them
4495 * to the array.
4496 * We currently require enough spares to make the final
4497 * array non-degraded. We also require that the difference
4498 * between old and new data_offset - on each device - is
4499 * enough that we never risk over-writing.
4500 */
4501
4502 unsigned long before_length, after_length;
4503 sector_t min_offset_diff = 0;
4504 int first = 1;
4505 struct geom new;
4506 struct r10conf *conf = mddev->private;
4507 struct md_rdev *rdev;
4508 int spares = 0;
4509 int ret;
4510
4511 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4512 return -EBUSY;
4513
4514 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4515 return -EINVAL;
4516
4517 before_length = ((1 << conf->prev.chunk_shift) *
4518 conf->prev.far_copies);
4519 after_length = ((1 << conf->geo.chunk_shift) *
4520 conf->geo.far_copies);
4521
4522 rdev_for_each(rdev, mddev) {
4523 if (!test_bit(In_sync, &rdev->flags)
4524 && !test_bit(Faulty, &rdev->flags))
4525 spares++;
4526 if (rdev->raid_disk >= 0) {
4527 long long diff = (rdev->new_data_offset
4528 - rdev->data_offset);
4529 if (!mddev->reshape_backwards)
4530 diff = -diff;
4531 if (diff < 0)
4532 diff = 0;
4533 if (first || diff < min_offset_diff)
4534 min_offset_diff = diff;
4535 first = 0;
4536 }
4537 }
4538
4539 if (max(before_length, after_length) > min_offset_diff)
4540 return -EINVAL;
4541
4542 if (spares < mddev->delta_disks)
4543 return -EINVAL;
4544
4545 conf->offset_diff = min_offset_diff;
4546 spin_lock_irq(&conf->device_lock);
4547 if (conf->mirrors_new) {
4548 memcpy(conf->mirrors_new, conf->mirrors,
4549 sizeof(struct raid10_info)*conf->prev.raid_disks);
4550 smp_mb();
4551 kfree(conf->mirrors_old);
4552 conf->mirrors_old = conf->mirrors;
4553 conf->mirrors = conf->mirrors_new;
4554 conf->mirrors_new = NULL;
4555 }
4556 setup_geo(&conf->geo, mddev, geo_start);
4557 smp_mb();
4558 if (mddev->reshape_backwards) {
4559 sector_t size = raid10_size(mddev, 0, 0);
4560 if (size < mddev->array_sectors) {
4561 spin_unlock_irq(&conf->device_lock);
4562 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4563 mdname(mddev));
4564 return -EINVAL;
4565 }
4566 mddev->resync_max_sectors = size;
4567 conf->reshape_progress = size;
4568 } else
4569 conf->reshape_progress = 0;
4570 conf->reshape_safe = conf->reshape_progress;
4571 spin_unlock_irq(&conf->device_lock);
4572
4573 if (mddev->delta_disks && mddev->bitmap) {
4574 struct mdp_superblock_1 *sb = NULL;
4575 sector_t oldsize, newsize;
4576
4577 oldsize = raid10_size(mddev, 0, 0);
4578 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4579
4580 if (!mddev_is_clustered(mddev)) {
4581 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4582 if (ret)
4583 goto abort;
4584 else
4585 goto out;
4586 }
4587
4588 rdev_for_each(rdev, mddev) {
4589 if (rdev->raid_disk > -1 &&
4590 !test_bit(Faulty, &rdev->flags))
4591 sb = page_address(rdev->sb_page);
4592 }
4593
4594 /*
4595 * some node is already performing reshape, and no need to
4596 * call md_bitmap_resize again since it should be called when
4597 * receiving BITMAP_RESIZE msg
4598 */
4599 if ((sb && (le32_to_cpu(sb->feature_map) &
4600 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4601 goto out;
4602
4603 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4604 if (ret)
4605 goto abort;
4606
4607 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4608 if (ret) {
4609 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4610 goto abort;
4611 }
4612 }
4613 out:
4614 if (mddev->delta_disks > 0) {
4615 rdev_for_each(rdev, mddev)
4616 if (rdev->raid_disk < 0 &&
4617 !test_bit(Faulty, &rdev->flags)) {
4618 if (raid10_add_disk(mddev, rdev) == 0) {
4619 if (rdev->raid_disk >=
4620 conf->prev.raid_disks)
4621 set_bit(In_sync, &rdev->flags);
4622 else
4623 rdev->recovery_offset = 0;
4624
4625 /* Failure here is OK */
4626 sysfs_link_rdev(mddev, rdev);
4627 }
4628 } else if (rdev->raid_disk >= conf->prev.raid_disks
4629 && !test_bit(Faulty, &rdev->flags)) {
4630 /* This is a spare that was manually added */
4631 set_bit(In_sync, &rdev->flags);
4632 }
4633 }
4634 /* When a reshape changes the number of devices,
4635 * ->degraded is measured against the larger of the
4636 * pre and post numbers.
4637 */
4638 spin_lock_irq(&conf->device_lock);
4639 mddev->degraded = calc_degraded(conf);
4640 spin_unlock_irq(&conf->device_lock);
4641 mddev->raid_disks = conf->geo.raid_disks;
4642 mddev->reshape_position = conf->reshape_progress;
4643 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4644
4645 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4646 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4647 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4648 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4649 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4650
4651 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4652 "reshape");
4653 if (!mddev->sync_thread) {
4654 ret = -EAGAIN;
4655 goto abort;
4656 }
4657 conf->reshape_checkpoint = jiffies;
4658 md_wakeup_thread(mddev->sync_thread);
4659 md_new_event();
4660 return 0;
4661
4662 abort:
4663 mddev->recovery = 0;
4664 spin_lock_irq(&conf->device_lock);
4665 conf->geo = conf->prev;
4666 mddev->raid_disks = conf->geo.raid_disks;
4667 rdev_for_each(rdev, mddev)
4668 rdev->new_data_offset = rdev->data_offset;
4669 smp_wmb();
4670 conf->reshape_progress = MaxSector;
4671 conf->reshape_safe = MaxSector;
4672 mddev->reshape_position = MaxSector;
4673 spin_unlock_irq(&conf->device_lock);
4674 return ret;
4675 }
4676
4677 /* Calculate the last device-address that could contain
4678 * any block from the chunk that includes the array-address 's'
4679 * and report the next address.
4680 * i.e. the address returned will be chunk-aligned and after
4681 * any data that is in the chunk containing 's'.
4682 */
last_dev_address(sector_t s,struct geom * geo)4683 static sector_t last_dev_address(sector_t s, struct geom *geo)
4684 {
4685 s = (s | geo->chunk_mask) + 1;
4686 s >>= geo->chunk_shift;
4687 s *= geo->near_copies;
4688 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4689 s *= geo->far_copies;
4690 s <<= geo->chunk_shift;
4691 return s;
4692 }
4693
4694 /* Calculate the first device-address that could contain
4695 * any block from the chunk that includes the array-address 's'.
4696 * This too will be the start of a chunk
4697 */
first_dev_address(sector_t s,struct geom * geo)4698 static sector_t first_dev_address(sector_t s, struct geom *geo)
4699 {
4700 s >>= geo->chunk_shift;
4701 s *= geo->near_copies;
4702 sector_div(s, geo->raid_disks);
4703 s *= geo->far_copies;
4704 s <<= geo->chunk_shift;
4705 return s;
4706 }
4707
reshape_request(struct mddev * mddev,sector_t sector_nr,int * skipped)4708 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4709 int *skipped)
4710 {
4711 /* We simply copy at most one chunk (smallest of old and new)
4712 * at a time, possibly less if that exceeds RESYNC_PAGES,
4713 * or we hit a bad block or something.
4714 * This might mean we pause for normal IO in the middle of
4715 * a chunk, but that is not a problem as mddev->reshape_position
4716 * can record any location.
4717 *
4718 * If we will want to write to a location that isn't
4719 * yet recorded as 'safe' (i.e. in metadata on disk) then
4720 * we need to flush all reshape requests and update the metadata.
4721 *
4722 * When reshaping forwards (e.g. to more devices), we interpret
4723 * 'safe' as the earliest block which might not have been copied
4724 * down yet. We divide this by previous stripe size and multiply
4725 * by previous stripe length to get lowest device offset that we
4726 * cannot write to yet.
4727 * We interpret 'sector_nr' as an address that we want to write to.
4728 * From this we use last_device_address() to find where we might
4729 * write to, and first_device_address on the 'safe' position.
4730 * If this 'next' write position is after the 'safe' position,
4731 * we must update the metadata to increase the 'safe' position.
4732 *
4733 * When reshaping backwards, we round in the opposite direction
4734 * and perform the reverse test: next write position must not be
4735 * less than current safe position.
4736 *
4737 * In all this the minimum difference in data offsets
4738 * (conf->offset_diff - always positive) allows a bit of slack,
4739 * so next can be after 'safe', but not by more than offset_diff
4740 *
4741 * We need to prepare all the bios here before we start any IO
4742 * to ensure the size we choose is acceptable to all devices.
4743 * The means one for each copy for write-out and an extra one for
4744 * read-in.
4745 * We store the read-in bio in ->master_bio and the others in
4746 * ->devs[x].bio and ->devs[x].repl_bio.
4747 */
4748 struct r10conf *conf = mddev->private;
4749 struct r10bio *r10_bio;
4750 sector_t next, safe, last;
4751 int max_sectors;
4752 int nr_sectors;
4753 int s;
4754 struct md_rdev *rdev;
4755 int need_flush = 0;
4756 struct bio *blist;
4757 struct bio *bio, *read_bio;
4758 int sectors_done = 0;
4759 struct page **pages;
4760
4761 if (sector_nr == 0) {
4762 /* If restarting in the middle, skip the initial sectors */
4763 if (mddev->reshape_backwards &&
4764 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4765 sector_nr = (raid10_size(mddev, 0, 0)
4766 - conf->reshape_progress);
4767 } else if (!mddev->reshape_backwards &&
4768 conf->reshape_progress > 0)
4769 sector_nr = conf->reshape_progress;
4770 if (sector_nr) {
4771 mddev->curr_resync_completed = sector_nr;
4772 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4773 *skipped = 1;
4774 return sector_nr;
4775 }
4776 }
4777
4778 /* We don't use sector_nr to track where we are up to
4779 * as that doesn't work well for ->reshape_backwards.
4780 * So just use ->reshape_progress.
4781 */
4782 if (mddev->reshape_backwards) {
4783 /* 'next' is the earliest device address that we might
4784 * write to for this chunk in the new layout
4785 */
4786 next = first_dev_address(conf->reshape_progress - 1,
4787 &conf->geo);
4788
4789 /* 'safe' is the last device address that we might read from
4790 * in the old layout after a restart
4791 */
4792 safe = last_dev_address(conf->reshape_safe - 1,
4793 &conf->prev);
4794
4795 if (next + conf->offset_diff < safe)
4796 need_flush = 1;
4797
4798 last = conf->reshape_progress - 1;
4799 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4800 & conf->prev.chunk_mask);
4801 if (sector_nr + RESYNC_SECTORS < last)
4802 sector_nr = last + 1 - RESYNC_SECTORS;
4803 } else {
4804 /* 'next' is after the last device address that we
4805 * might write to for this chunk in the new layout
4806 */
4807 next = last_dev_address(conf->reshape_progress, &conf->geo);
4808
4809 /* 'safe' is the earliest device address that we might
4810 * read from in the old layout after a restart
4811 */
4812 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4813
4814 /* Need to update metadata if 'next' might be beyond 'safe'
4815 * as that would possibly corrupt data
4816 */
4817 if (next > safe + conf->offset_diff)
4818 need_flush = 1;
4819
4820 sector_nr = conf->reshape_progress;
4821 last = sector_nr | (conf->geo.chunk_mask
4822 & conf->prev.chunk_mask);
4823
4824 if (sector_nr + RESYNC_SECTORS <= last)
4825 last = sector_nr + RESYNC_SECTORS - 1;
4826 }
4827
4828 if (need_flush ||
4829 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4830 /* Need to update reshape_position in metadata */
4831 wait_barrier(conf, false);
4832 mddev->reshape_position = conf->reshape_progress;
4833 if (mddev->reshape_backwards)
4834 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4835 - conf->reshape_progress;
4836 else
4837 mddev->curr_resync_completed = conf->reshape_progress;
4838 conf->reshape_checkpoint = jiffies;
4839 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4840 md_wakeup_thread(mddev->thread);
4841 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4842 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4843 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4844 allow_barrier(conf);
4845 return sectors_done;
4846 }
4847 conf->reshape_safe = mddev->reshape_position;
4848 allow_barrier(conf);
4849 }
4850
4851 raise_barrier(conf, 0);
4852 read_more:
4853 /* Now schedule reads for blocks from sector_nr to last */
4854 r10_bio = raid10_alloc_init_r10buf(conf);
4855 r10_bio->state = 0;
4856 raise_barrier(conf, 1);
4857 atomic_set(&r10_bio->remaining, 0);
4858 r10_bio->mddev = mddev;
4859 r10_bio->sector = sector_nr;
4860 set_bit(R10BIO_IsReshape, &r10_bio->state);
4861 r10_bio->sectors = last - sector_nr + 1;
4862 rdev = read_balance(conf, r10_bio, &max_sectors);
4863 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4864
4865 if (!rdev) {
4866 /* Cannot read from here, so need to record bad blocks
4867 * on all the target devices.
4868 */
4869 // FIXME
4870 mempool_free(r10_bio, &conf->r10buf_pool);
4871 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4872 return sectors_done;
4873 }
4874
4875 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4876 GFP_KERNEL, &mddev->bio_set);
4877 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4878 + rdev->data_offset);
4879 read_bio->bi_private = r10_bio;
4880 read_bio->bi_end_io = end_reshape_read;
4881 r10_bio->master_bio = read_bio;
4882 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4883
4884 /*
4885 * Broadcast RESYNC message to other nodes, so all nodes would not
4886 * write to the region to avoid conflict.
4887 */
4888 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4889 struct mdp_superblock_1 *sb = NULL;
4890 int sb_reshape_pos = 0;
4891
4892 conf->cluster_sync_low = sector_nr;
4893 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4894 sb = page_address(rdev->sb_page);
4895 if (sb) {
4896 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4897 /*
4898 * Set cluster_sync_low again if next address for array
4899 * reshape is less than cluster_sync_low. Since we can't
4900 * update cluster_sync_low until it has finished reshape.
4901 */
4902 if (sb_reshape_pos < conf->cluster_sync_low)
4903 conf->cluster_sync_low = sb_reshape_pos;
4904 }
4905
4906 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4907 conf->cluster_sync_high);
4908 }
4909
4910 /* Now find the locations in the new layout */
4911 __raid10_find_phys(&conf->geo, r10_bio);
4912
4913 blist = read_bio;
4914 read_bio->bi_next = NULL;
4915
4916 rcu_read_lock();
4917 for (s = 0; s < conf->copies*2; s++) {
4918 struct bio *b;
4919 int d = r10_bio->devs[s/2].devnum;
4920 struct md_rdev *rdev2;
4921 if (s&1) {
4922 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4923 b = r10_bio->devs[s/2].repl_bio;
4924 } else {
4925 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4926 b = r10_bio->devs[s/2].bio;
4927 }
4928 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4929 continue;
4930
4931 bio_set_dev(b, rdev2->bdev);
4932 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4933 rdev2->new_data_offset;
4934 b->bi_end_io = end_reshape_write;
4935 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4936 b->bi_next = blist;
4937 blist = b;
4938 }
4939
4940 /* Now add as many pages as possible to all of these bios. */
4941
4942 nr_sectors = 0;
4943 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4944 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4945 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4946 int len = (max_sectors - s) << 9;
4947 if (len > PAGE_SIZE)
4948 len = PAGE_SIZE;
4949 for (bio = blist; bio ; bio = bio->bi_next) {
4950 /*
4951 * won't fail because the vec table is big enough
4952 * to hold all these pages
4953 */
4954 bio_add_page(bio, page, len, 0);
4955 }
4956 sector_nr += len >> 9;
4957 nr_sectors += len >> 9;
4958 }
4959 rcu_read_unlock();
4960 r10_bio->sectors = nr_sectors;
4961
4962 /* Now submit the read */
4963 md_sync_acct_bio(read_bio, r10_bio->sectors);
4964 atomic_inc(&r10_bio->remaining);
4965 read_bio->bi_next = NULL;
4966 submit_bio_noacct(read_bio);
4967 sectors_done += nr_sectors;
4968 if (sector_nr <= last)
4969 goto read_more;
4970
4971 lower_barrier(conf);
4972
4973 /* Now that we have done the whole section we can
4974 * update reshape_progress
4975 */
4976 if (mddev->reshape_backwards)
4977 conf->reshape_progress -= sectors_done;
4978 else
4979 conf->reshape_progress += sectors_done;
4980
4981 return sectors_done;
4982 }
4983
4984 static void end_reshape_request(struct r10bio *r10_bio);
4985 static int handle_reshape_read_error(struct mddev *mddev,
4986 struct r10bio *r10_bio);
reshape_request_write(struct mddev * mddev,struct r10bio * r10_bio)4987 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4988 {
4989 /* Reshape read completed. Hopefully we have a block
4990 * to write out.
4991 * If we got a read error then we do sync 1-page reads from
4992 * elsewhere until we find the data - or give up.
4993 */
4994 struct r10conf *conf = mddev->private;
4995 int s;
4996
4997 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4998 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4999 /* Reshape has been aborted */
5000 md_done_sync(mddev, r10_bio->sectors, 0);
5001 return;
5002 }
5003
5004 /* We definitely have the data in the pages, schedule the
5005 * writes.
5006 */
5007 atomic_set(&r10_bio->remaining, 1);
5008 for (s = 0; s < conf->copies*2; s++) {
5009 struct bio *b;
5010 int d = r10_bio->devs[s/2].devnum;
5011 struct md_rdev *rdev;
5012 rcu_read_lock();
5013 if (s&1) {
5014 rdev = rcu_dereference(conf->mirrors[d].replacement);
5015 b = r10_bio->devs[s/2].repl_bio;
5016 } else {
5017 rdev = rcu_dereference(conf->mirrors[d].rdev);
5018 b = r10_bio->devs[s/2].bio;
5019 }
5020 if (!rdev || test_bit(Faulty, &rdev->flags)) {
5021 rcu_read_unlock();
5022 continue;
5023 }
5024 atomic_inc(&rdev->nr_pending);
5025 rcu_read_unlock();
5026 md_sync_acct_bio(b, r10_bio->sectors);
5027 atomic_inc(&r10_bio->remaining);
5028 b->bi_next = NULL;
5029 submit_bio_noacct(b);
5030 }
5031 end_reshape_request(r10_bio);
5032 }
5033
end_reshape(struct r10conf * conf)5034 static void end_reshape(struct r10conf *conf)
5035 {
5036 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
5037 return;
5038
5039 spin_lock_irq(&conf->device_lock);
5040 conf->prev = conf->geo;
5041 md_finish_reshape(conf->mddev);
5042 smp_wmb();
5043 conf->reshape_progress = MaxSector;
5044 conf->reshape_safe = MaxSector;
5045 spin_unlock_irq(&conf->device_lock);
5046
5047 if (conf->mddev->queue)
5048 raid10_set_io_opt(conf);
5049 conf->fullsync = 0;
5050 }
5051
raid10_update_reshape_pos(struct mddev * mddev)5052 static void raid10_update_reshape_pos(struct mddev *mddev)
5053 {
5054 struct r10conf *conf = mddev->private;
5055 sector_t lo, hi;
5056
5057 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
5058 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
5059 || mddev->reshape_position == MaxSector)
5060 conf->reshape_progress = mddev->reshape_position;
5061 else
5062 WARN_ON_ONCE(1);
5063 }
5064
handle_reshape_read_error(struct mddev * mddev,struct r10bio * r10_bio)5065 static int handle_reshape_read_error(struct mddev *mddev,
5066 struct r10bio *r10_bio)
5067 {
5068 /* Use sync reads to get the blocks from somewhere else */
5069 int sectors = r10_bio->sectors;
5070 struct r10conf *conf = mddev->private;
5071 struct r10bio *r10b;
5072 int slot = 0;
5073 int idx = 0;
5074 struct page **pages;
5075
5076 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5077 if (!r10b) {
5078 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5079 return -ENOMEM;
5080 }
5081
5082 /* reshape IOs share pages from .devs[0].bio */
5083 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5084
5085 r10b->sector = r10_bio->sector;
5086 __raid10_find_phys(&conf->prev, r10b);
5087
5088 while (sectors) {
5089 int s = sectors;
5090 int success = 0;
5091 int first_slot = slot;
5092
5093 if (s > (PAGE_SIZE >> 9))
5094 s = PAGE_SIZE >> 9;
5095
5096 rcu_read_lock();
5097 while (!success) {
5098 int d = r10b->devs[slot].devnum;
5099 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5100 sector_t addr;
5101 if (rdev == NULL ||
5102 test_bit(Faulty, &rdev->flags) ||
5103 !test_bit(In_sync, &rdev->flags))
5104 goto failed;
5105
5106 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5107 atomic_inc(&rdev->nr_pending);
5108 rcu_read_unlock();
5109 success = sync_page_io(rdev,
5110 addr,
5111 s << 9,
5112 pages[idx],
5113 REQ_OP_READ, 0, false);
5114 rdev_dec_pending(rdev, mddev);
5115 rcu_read_lock();
5116 if (success)
5117 break;
5118 failed:
5119 slot++;
5120 if (slot >= conf->copies)
5121 slot = 0;
5122 if (slot == first_slot)
5123 break;
5124 }
5125 rcu_read_unlock();
5126 if (!success) {
5127 /* couldn't read this block, must give up */
5128 set_bit(MD_RECOVERY_INTR,
5129 &mddev->recovery);
5130 kfree(r10b);
5131 return -EIO;
5132 }
5133 sectors -= s;
5134 idx++;
5135 }
5136 kfree(r10b);
5137 return 0;
5138 }
5139
end_reshape_write(struct bio * bio)5140 static void end_reshape_write(struct bio *bio)
5141 {
5142 struct r10bio *r10_bio = get_resync_r10bio(bio);
5143 struct mddev *mddev = r10_bio->mddev;
5144 struct r10conf *conf = mddev->private;
5145 int d;
5146 int slot;
5147 int repl;
5148 struct md_rdev *rdev = NULL;
5149
5150 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5151 if (repl)
5152 rdev = conf->mirrors[d].replacement;
5153 if (!rdev) {
5154 smp_mb();
5155 rdev = conf->mirrors[d].rdev;
5156 }
5157
5158 if (bio->bi_status) {
5159 /* FIXME should record badblock */
5160 md_error(mddev, rdev);
5161 }
5162
5163 rdev_dec_pending(rdev, mddev);
5164 end_reshape_request(r10_bio);
5165 }
5166
end_reshape_request(struct r10bio * r10_bio)5167 static void end_reshape_request(struct r10bio *r10_bio)
5168 {
5169 if (!atomic_dec_and_test(&r10_bio->remaining))
5170 return;
5171 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5172 bio_put(r10_bio->master_bio);
5173 put_buf(r10_bio);
5174 }
5175
raid10_finish_reshape(struct mddev * mddev)5176 static void raid10_finish_reshape(struct mddev *mddev)
5177 {
5178 struct r10conf *conf = mddev->private;
5179
5180 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5181 return;
5182
5183 if (mddev->delta_disks > 0) {
5184 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5185 mddev->recovery_cp = mddev->resync_max_sectors;
5186 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5187 }
5188 mddev->resync_max_sectors = mddev->array_sectors;
5189 } else {
5190 int d;
5191 rcu_read_lock();
5192 for (d = conf->geo.raid_disks ;
5193 d < conf->geo.raid_disks - mddev->delta_disks;
5194 d++) {
5195 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5196 if (rdev)
5197 clear_bit(In_sync, &rdev->flags);
5198 rdev = rcu_dereference(conf->mirrors[d].replacement);
5199 if (rdev)
5200 clear_bit(In_sync, &rdev->flags);
5201 }
5202 rcu_read_unlock();
5203 }
5204 mddev->layout = mddev->new_layout;
5205 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5206 mddev->reshape_position = MaxSector;
5207 mddev->delta_disks = 0;
5208 mddev->reshape_backwards = 0;
5209 }
5210
5211 static struct md_personality raid10_personality =
5212 {
5213 .name = "raid10",
5214 .level = 10,
5215 .owner = THIS_MODULE,
5216 .make_request = raid10_make_request,
5217 .run = raid10_run,
5218 .free = raid10_free,
5219 .status = raid10_status,
5220 .error_handler = raid10_error,
5221 .hot_add_disk = raid10_add_disk,
5222 .hot_remove_disk= raid10_remove_disk,
5223 .spare_active = raid10_spare_active,
5224 .sync_request = raid10_sync_request,
5225 .quiesce = raid10_quiesce,
5226 .size = raid10_size,
5227 .resize = raid10_resize,
5228 .takeover = raid10_takeover,
5229 .check_reshape = raid10_check_reshape,
5230 .start_reshape = raid10_start_reshape,
5231 .finish_reshape = raid10_finish_reshape,
5232 .update_reshape_pos = raid10_update_reshape_pos,
5233 };
5234
raid_init(void)5235 static int __init raid_init(void)
5236 {
5237 return register_md_personality(&raid10_personality);
5238 }
5239
raid_exit(void)5240 static void raid_exit(void)
5241 {
5242 unregister_md_personality(&raid10_personality);
5243 }
5244
5245 module_init(raid_init);
5246 module_exit(raid_exit);
5247 MODULE_LICENSE("GPL");
5248 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5249 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5250 MODULE_ALIAS("md-raid10");
5251 MODULE_ALIAS("md-level-10");
5252